TRANSPORTATION RESEARCH RECORD

No. 1506 Aviation

Airport and Air Transportation Issues

A peer-reviewed publication of the Transportation Research Board

TRANSPORTATION RESEARCH BOARD NATIONAL RESEARCH COUNCIL

NATIONAL ACADEMY PRESS WASHINGTON, D.C. 1995 Transportation Research Record 1506 Sponsorship of Transportation Research Record 1506 ISSN 0361-1981 ISBN 0-309-06168-7 GROUP I-TRANSPORTATION SYSTEMS PLANNING AND Price: $25.00 ADMINISTRATION Chairman: Thomas F. Humphrey, Massachusetts Institute of Technology Subscriber Category V aviation A via ti on Section Chairman: John W. Fischer, Library of Congress Printed in the United States of America Committee on Aviation Economics and Forecasting Chairman: Vicki L. Golich, California State University-San Marcos

Committee on Airport Terminals and Ground Access Chairman: Andrew C. Lerner, The Matrix Group Inc. Secretary: Peter B. Mandie, Leigh Fisher Associates John S. Andrews, Zale Anis, Claire Barrett, Winfield S. Beyea, William E. Brougher, Jody Yamanaka Clovis, James A. Doyle, Norman J. Faramelli, Sa/ah G. Hamzawi, Dean Hobson, Bradley T. Jacobsen, M. Julie Kim, Saleh A. Mumayiz, Edward G. O'Rourke, Siegbert B. Poritzky, Panos D. Prevedouros, Prianka N. Seneviratne, Vojin Tosic, S. C. Wirasinghe

Committee on Airfield and Airspace Capacity and Delay Chairman: Agam N. Sinha, The Mitre Corporation Secretary: Jeff S. Mishler, HNTB Corporation Gregory A. Casto, Melvyn Ches/ow, Lawrence F. Cunningham, Ray H. Fowler, Evan C. Futterman, Douglas F. Goldberg, Belinda G. Hargrove, Jay Etta Z. Hecker, John E. Lebron, Andrew B. Levy, Michael T. McNerney, James P. Muldoon, Saleh A. Mumayiz, Clyde W. Pace, Panos D. Prevedouros, David J. Shefte/, Edmund Spring, Juan F. Tapia-Videla, Nicholas G. Tsongos, Alan Yazdani

Transportation Research Board Staff Robert E. Spicher, Director, Technical Activities Joseph A. Breen, Aviation Specialist Nancy A. Ackerman, Director, Reports and Editorial Services

Sponsorship is indicated by a footnote at the end of each paper. The organi­ zational units, officers, and members are as of December 31, 1994. 1 Transportation Research Record 1506

Contents

Foreword v

The Future of Business Air Travel 1 David E. Raphael and Claire Starry

Multiple Airport Systems in the United States: Current Status 8 and Future Prospects Mark Hansen and Tara Weidner

Forecasting Air Travel Arrivals: Model Development and Application 18 at the Honolulu International Airport Sanjay Kawad and Panos D. Prevedouros

Flight Sequencing in Airport Hub Operations 26 Ching Change and Paul Schonfeld

An Optimum Resource Utilization Plan for Airport Passenger 34 Terminal Buildings Mahmoud S. Parizi and John P. Braaksma

Analysis of Moving Walkway Use in Airport Terminal Corridors 44 Seth Young

Characterization of Gate Location on Aircraft Runway Landing Roll 52 Prediction and Airport Ground Networks Navigation Xiaoling Gu, Antonio A. Trani, and Caoyuan Zhong

Economic Characteristics of Multiple Vehicle Delivery Tours Satisfying 61 Time Constraints Max K. Kiesling and Mark M. Hansen

Challenges in Developing an Airport Empolyee Commute Program: 70 Case Study of Boston Logan International Airport Diane M. Ricard

Foreword

Eight of the nine papers in this volume focus on various aspects of modern airports or airport systems, and their associated passenger, cargo, and aircraft activity. The remaining paper, Raphael and Starry's The Future ofBusiness Air Travel, examines factors that may portend changes in the traditional pattern and frequency of business air travel. Hansen and Weidner's paper, Multiple Airport Systems in the United States: Current Status and Future Prospects, reviews arid assesses the prospects for new Multiple Airport Systems (MASs) in the United States and concludes that the U. S. air transport system has reached a point in which MASs could become a competitive industry in many regions.· Hawaii's major and secondary airports are the subject of Kawad and Prevedouros' paper, Forecasting Air Travel Arrivals: Model Development and Application at the Honolulu International Airport, in which the authors develop a short-to-medium-term econometric model system for forecasting air traffic arrivals. In their paper on Flight Sequencing in Airport Hub Operations, Chang and Schonfeld propose an efficient heuristic method to minimize the total costs of passenger transfer, aircraft ground time, and gate utilization. Parizi and Braaksma's paper, An Optimum Resource Utilization Plan for Airport Passenger Termi­ nal Buildings, describes an operational planning tool that, with additional research, has the potential to be used in operating Passenger Facility Buildings (PFBs) at their minimum over- and under-supply cost. The topic of pedestrian conveyor systems is analyzed in a paper by Young (Analysis of Moving Walkway Use in Airport Terminal Corridors). Young compares these systems with other primary air . terminal modes and speculates how altering certain characteristics could increase the conveyer sys­ tems' overall utility. Gu, Trani, and Zhong, in Characterization of Gate Location on Aircraft Runway Landing Roll Pre­ diction and fiirport Ground Networks Navigation, present and aircraft landing simulation and predic­ tion model using simple kinematics, coupled with individual parameters to describe the landing process. The model could help solve ground network traffic congestion problems and improve safety. Kiesling and Hansen note, in Economic Characteristics ofMultiple Vehicle Delivery, that, compared with the body of literature on passenger carrier operations, there is a gap in addressing the economics of air freight transportation. Their paper helps fill this gap by addressing the economic structure of ground-side freight distribution for air express carriers. Ricard's Challenges in Developing an Airport Employee Commute Program: Logan International Airport characterizes the commuting patterns of airport employees and describes available alternatives. The paper reviews the effectiveness of these alternatives, and summarizes past and future initiatives to alter current patterns.

v

TRANSPORTATION RESEARCH RECORD 1506

The Future of Business Air Travel

DAVIDE. RAPHAEL AND CLAIRE STARRY

Recent data indicate that business travel is slowing, and most business To help understand the extent to which the above and other fac­ travelers are aggressively pursuing travel management policies that tors are influencing business travel, this study attempts to answer include limits on travel and negotiations with airlines for lower fares. the following questions: Some analysts find that business air travel may be further adversely affected from the proliferation of communications technologies, includ­ ing teleconferencing. Four findings are discussed: Statistically signifi­ 1. Has there been a downward shift in the elasticity of demand cant changes in the relationship of business air travel to gross domestic for air travel with respect to gross domestic product? product (GDP) occurred in the late 1980s; recovery of business travel 2. Which industrial sectors are the major purchasers of business is likely to be less robust compared with previous business cycles. air travel, and what is likely to be happening to their growth and Econometric analysis of business and total passenger enplanements in employment profile over the next several years? the U.S. domestic air system indicate that a significant decline in the 3. Are travel management policies changing corporate air travel? elasticity of demand with respect to GDP occurred in the late 1980s and early 1990s. About 40 industrial sectors account for 80 percent of busi­ 4. How will teleconferencing and other communications tech­ ness air travel. Median job and output growth for many of these sectors nologies affect business travel? are below the national average. Many companies now turn to travel managers or travel service organizations and third-party firms to man­ This report is limited to evaluating U.S. domestic passenger age travel. Business air travelers are no longer willing to pay substan­ enplanements, and relies heavily on secondary data from the U.S. tially higher air fares than personal travelers pay and have the skills to Department of Transportation, Federal Aviation Administration counter the airlines' yield management programs. Most industries are familiar with telecommunications technologies and anecdotal evidence (FAA) (4), U.S. Department of Commerce, Bureau of Economic indicates many companies are currently substituting teleconferences for Analysis (BEA) (5), and the U.S. _Department of Labor, Bureau of travel, at least for intracompany meetings. Future advances in telecom­ Labor Statistics (BLS) (6). munications and electronic communications offer additional convenient and less expensive alternatives to air travel. THE ELASTICITY OF DEMAND FOR AIR TRAVEL

During the 1970s and 1980s, business travel growth was an impor­ Historically, most analysts found the elasticity of demand for busi­ tant component of airline industry profitability and performance. ness travel to be inelastic with respect to yields (or fares) and elas­ Business travel was the mainstay of the industry, providing suffi­ tic with respect to an income variable, most commonly real gross cient yields to more than cover costs and enabling airlines to offer domestic product (GDP). Since 1988, growth rates in total enplane­ substantial discounts to personal travelers. Recent data indicate that ments have slowed. Figure 1 shows the actual growth in total pas­ business travel may no longer be growing, and more business trav­ senger enplanements compared with the predicted growth using a elers are taking advantage of discount fares. For example, in 1981, regression equation based on 1969-1987 data ( 4). The gap that discount revenue passenger miles (RPMs) accounted for 70 percent starts in 1988 reaches about 10 percent by 1993 and declines to of the total; by 1991, this percentage had risen to over 95 percent about 6% in 1994. (If log linear equations are used, the difference (1). Several reasons for the change in business travel have been sug­ between forecast and actual is over 20%.) Most of the 1994 growth gested, including: in enplanements, however, was accounted for by short haul trips and discount fares. Data from passenger surveys (3) suggest that • Teleconferencing is beginning to take off. One expert pre­ business enplanements as a percentage of total enplanements dicts that telecommunications will substitute for 25 percent of busi­ declined about 5 percentage points over the late 1980s and early ness travel by 2010 (2). Another survey indicates a small 1990s. percentage of business air travel has already been diverted to To test the hypothesis that the relationship between business air telecommunications (3). travel and GDP has changed, we conducted an econometric analy­ • Businesses recognize that air travel costs are substantial and sis using the regression equations shown in Table 1. From 1988 take steps to control these costs through use of travel management onward, a dummy variable is used. If this variable is significantly policies. Many companies are imposing policies that lower fares different from zero (a t static greater than 1.8 at the 95 percent sig­ and impose travel restrictions. nifi_cance level), the hypothesis of a change in the relationship is not • Corporate downsizing, especially middle managers, likely will rejected. For both the linear and log-linear models, the dummy vari­ result in fewer business travelers. able is significant, indicating that such a change in the business air travel elasticity of demand has occurred. Time series data for more years are available for total passenger enplanements. Additional tests are made using these data, and they D. E. Raphael, Marcar Management Institute, 4007 Marsten Avenue, Bel­ also support the hypothesis that significant changes in the demand mont, Calif. 94002. C. Starry, TDS Economics, 1040 Noel Drive, Suite 202, elasticity relationships occurred at the end of the 1980s. Table 2 pre­ Menlo Park, Calif. 94025. sents the estimation results for linear and log-linear models that com- 2 TRANSPORTATION RESEARCH RECORD 1506

550,000 500,000 . 450,000 / C/) 'E ..... ,..-'·-· Q) -· 400,000 E Q) c ro a.. 350,000 wc 0 CJ) 300,000 "'O c: ro C/) :l 250,000 0 --•--Actual ..._..t::. 200,000 ---0- Predicted

150,000

100,000

FIGURE 1 Actual versus predicted total domestic passenger enplanements: 1969-1994. pare the 1969-1987 period to the 1988-1994 period. A test to deter­ Preliminary data are uncertain about the change in pas­ mine if the coefficients are significantly different (7) is conducted for senger enplanements in 1994. Surveys of business travelers sug­ both models. The results, summarized in Table 3, indicate a differ­ gest business travel will be up, but often travel budgets are the ence in coefficients at the 95 percent level for both the linear and log­ same or lower. Many companies report that they will be increasing linear models, and a difference in coefficients at the 99 percent level the number of airline trips but spending less on air fares because for the log-linear model. We also tested the hypothesis using dummy they have implemented cost control and travel management variables for the 1988-1994 period. For both the linear and log­ policies (8). linear models, the dummy variables are significant at the 95 percent level, and at the 99 percent level for the log-linear model. Econometric analysis definitely supports the contention that a INDUSTRY SPECIFIC ANALYSIS major change has occurred in the functional relationship between business travel and traditional explanatory variables, GDP, and Major Purchasers of Air Transportation yield. This shift indicates that business enplanements, while con­ tinuing to grow with the economy, will be growing at a rate lower As shown in Figure 2, the manufacturing industries account for less than that of GDP or similar variable. than 20 percent of business air travel expenditures. Federal, state,

TABLE 1 Estimated Elasticities of Demand for Business Air Travel, 1977-1992

Equation R2 F-Statistic GDP Elasticity Yield Elasticity Dummy Variablet

Linear .985 99.78 1.53 -0.13 -276.26 (6.4) (-1.l) (-2.0) Log Linear .982 79.16 1.65 -0.18 -0.07 (6.0) (-0.9) (-2.0)

t Set equal to zero for years 1977 to 1988, and equal to 1for1989 to 1992. Note: t-statistics given in parenthesis. Raphael and Stany 3

TABLE2 Estimated Elasticities of Demand for Total Air Travel, 1969-1994 GDP Yield Dummy Time Period Adjusted R2 F Statistic Elasticity Elasticity Variablet

Linear Equations 1988-1994 0.41 3.06 0.61 -0.21 (1.12) (-0.50) 1969-1987 0.98 384.76 1.25 -0.32 (8.18) (-2.99) 1969-1994 0.98 648.83 1.03 -0.39 (6.66) (-3.07) 1969-1994 0.98 450.87 1.19 -0.34 -20,097 (6.92) (-2.80) (-1.79)

Log-Linear Equations 1988-1994 0.40 3.04 0.58 -0.23 (1.03) (-0.59) 1969-1987 0.98 559.33 1.87 -0.41 (12.83) (-2.47) 1969-1994 0.97 451.49 1.76 -0.19 (8.32) (-0.78) 1969-1994 0.98 495.66 1.91 -0.31 -0.14 (11.22) (-1.62) (-3.93) t Dummy variable equals O for 1969 to 1987, and 1 for 1988 to 1993. A significant value indicates that an adjustment to the equation occurred during the second time period. Note: t-statistics given in parenthesis.

and local government sectors account for about 12% of business air median growth rate in employment for these industries is 1.0 per­ travel. The remaining two-thirds of business air travel is primarily cent and the average growth in real output is 2.1 percent. These in the services sectors, including wholesale and retail trade, finance, compare with 1.3 and 2.7 percent, respectively, for the U.S. econ­ insurance, and real estate, and a broad range of other services such omy. Only 10 of the 40 top air travel industries are forecast to have as management consulting, legal, medical, and educational services. faster job growth between 1992 and 2005 compared with their job Telecommunications expenditures (discussed in the section below) growth during the period 1979-1992. are more highly skewed toward the communications (about a third The top 10 business air travel sectors include the U.S. Depart­ of expenditures on communications come from other communica­ ment of Defense, which is undergoing cutbacks in funding and per­ tions industries or companies), and are less likely to be affected by sonnel; the U.S. Post Office, which is faced with increasing com­ the downturn in aerospace and 9efense and declines in the number petition from the private sector, E-mail and facsimile machines; and of manufacturing jobs. public education, which is suffering from financial problems in Using BEA data (4), we identified the top 40 out of approxi­ many parts of the country. Several of the top sectors are considered mately 480 industrial sectors that accounted for 80 percent of busi- to be high growth ones, at least in terms of employment. These . ness expenditures on business air travel in the late 1980s. (See include management consulting and retail trade. High growth man­ Table 4.) This section evaluates these industries in terms of their ufacturing sectors, including semiconductors and computers, are past and future growth prospects, because to a considerable extent, also among the top industries. Although some growth in business the fortunes of these 40 sectors dictate the fortunes of the air air travel is expected from the manufacturing sectors, the emphasis industry. placed on cost control could easily keep expenditures level, con­ tinue downward pressure on yields, and moderate the growth in enplanements. Growth Prospects Overall, the composition of industries that account for 80 percent of business air travel is changing, and future shifts will emphasize As a whole, the top 40 sectors are forecast by the BLS (5) to grow the trend toward the service sectors being the primary business air slower than the national average. For the decade of the 1990s, the travel users.

TABLE 3 Test for Equality between Coefficients Model FRaho Reject Null Reject Null Hypothesis at 5 % Hypothesis at 1 % Significance Levelt Significance Levelt

Linear 3.80 Yes No Log linear 13.06 Yes Yes

t Null hypothesis is that the coefficients for the 1969 to 1987 period are equal (not significantly different from) the coefficients for the 1988 to 1993 period. For 5%, the F ratio must exceed 3.10 to reject the null hypothesis; for 1% the F ratio must exceed 4.94. 4 TRANSPORTATION RESEARCH RECORD 1506

0 45% Q) > .=cu 40% :.a: 35% c: • Passenger Air 0 (/) (/) c: 30% Q) 0 -~ Communications ~ (.) El :-0 25% c: ·c: Q) :i a. E x w E 20% (/) 8 (/) Q) Q) Q) 15% c "(;) ~ :i CXl 10% 0 c: 5% Q) (.) Qi a.. 0% c O'l vi Q) (/) w Q) c: 0 c c c "'O Q) i- a: (.) -~ 0 0 iii o, ·u (/) ~ E Q) -~ -~ u::: -~ c Q) ~ c 2 t5 (.) Q) 0 ·c: cu t Q; u... iii 0 ·c: ~ Cf') >. ~ c "S a. :i o, 0 c (/) 5 0 (.) cu c: E (!) <( cu E ~ 0 .= (.)

FIGURE 2 Distribution of air travel and telecommunications expenditures by sector.

USE OF TRAVEL MANAGERS looks for ways to control costs. At the other extreme, consulting, legal, some financial, and similar industries are lax in implementing There has been a continued increase in the number of businesses and enforcing policies and are not aggressive in seeking out the low­ using travel agents and travel managers to help control costs. The est possible air fare. The type of trip most likely to be restricted is percentage of corporations that have corporate travel managers has that involved with meeting other employees of the same company. grown from 7.5 to 38% between 1982 and 1993 (9). Nearly 80% of Conversely, the type of trip least likely to be restricted is one involv­ ·companies surveyed that spend more than $5 million on travel and ing sales calls to existing or prospective clients. entertainment (T &E) have such a manager. The role of travel man­ ager has become more important as companies take steps to control travel and formalize policies. These managers generally report to a COMPETITION FROM TELECOMMUNICATIONS corporate administrative office. Data from the 1991 Travel Weekly survey indicate that 23% of All industries use telecommunications, and telecommunications corporate clients of travel agencies had written guidelines or poli­ expenditures generally exceed those of air transportation by about cies that year. By 1994, this percentage increased to 38%. The major 2.5 to 1. The same is true for industrial sectors that are the main­ focus of the policies is on air travel. The percentage of agency book­ stay of business air travel. (See Figure 3.) Many sectors that offer ings coming from corporate travel managers or coordinators the greatest prospects for growth, however, also tend to be those increased from 9% in 1991 to 12% in 1994 (10). The American that have the highest ratio of telecommunications expenditures to Express survey found that 64% of companies with 100 employees air travel expenditures: for example, trade, transportation, elec­ or more had formal written travel guidelines. Large companies are tronics and computers, and communications equipment and some much more likely to have policies: 96% of companies with T &E business services. These industries are among those most likely to budgets over $5 million and 90% of companies with budgets from embrace telecommunications as an accepted means of doing $1 million to $5 million report that they had formal written guide­ business. lines in place, compared with only 28% of companies with T&E Other studies support the growing use of telecommunications. budgets under $100,000. · One indicates the extent of substitution should be 5 percent by 2000 Variation is wide among companies and industries in terms of and 25 percent by 2010 (2). Clearly substitution will depend on the travel policies, cost control, and use of telecommunications. Com­ availability and cost of telecommunications versus air travel, and panies in manufacturing and engineering industries are generally also the acceptance of telecommunications compared with face-to­ the most cost conscious. Any company, however, that operates in a face contact. Telecommunications costs will continue to decline rel­ very competitive environment carefully scrutinizes air travel and ative to air travel costs. The ability of communications to replace Raphael and Starry 5

TABLE4 Estimated Expenditures on Air Travel by Selected Industries, 1977, 1982, and 1987 1977 1982 1987 1982-87 Growth rate Millions of Dollars Wholesale trade 1,038 1,950 4,307 17% Air transportationt 867 1,396 4,113 24% Consulting services 316 773 2,641 28% State and local government 785 1,210 1,866 9% Federal government, defense 614 1,484 1,817 4% Security and commodity brokers 124 647 1,243 14% Associations 9 560 1, 141 15% Legal services 184 248 1,097 35% U.S. Postal Service 460 695 906 5% Retail trade 325 707 763 2% Colleges, universities 187 291 664 18% Federal government, nondefense 279 422 648 9% Banking 90 181 599 27% Aircraft 1,593 435 569 6% Doctors and dentists 84 164 509 25% Freight forwarders and other transportation services 98 200 483 19% Periodicals 87 199 467 19% Newspapers 225 500 445 -2% Construction 222 296 433 8% Real estate 98 219 418 14% Electronic computers 121 481 341 -7% Electric services (utilities) 96 229 325 7% Engineering services 196 244 310 5% Other business services 44 88 305 28% Insurance carriers 130 283 288 0% Other membership organizations 60 212 283 6% Computer and data processing services 121 112 263 19% Motor freight transportation and warehousing 56 118 232 14% Accounting, auditing and bookkeeping, and 93 67 229 28% Insurance agents, brokers, and services 47 91 205 18% Radio and TV broadcasting 49 95 200 16% Credit agencies other than banks 53 124 191 9% Labor organizations, civic, social, and fraternal 58 98 184 13% Communications except radio and TV 50 122 183 8% Motor vehicles and car bodies 106 295 177 -10% Industrial inorganic and organic chemicals 129 118 170 7% Computer peripheral equipment 166 Aircraft and missile engines and engine parts 77 147 165 2% Agricultural, forestry, and fishery services 71 137 164 4% Commercial printing 60 ill 143 5% Total business and government 12,163 21,912 35,526 10% Personal consumption 12,769 20,574 29,349 7% Exports 2,146 5,691 10,186 12% Imports -2,233 -3,931 -5,711 8% Total industry output 24,846 44,245 69,350 9% t Air Transportation includes imputed value of services provided by one airline to another. Source: U.S. Deparuncnt of Commerce, Bureau of Economic Analysis.

face-to-face meetings may depend, to a large measure, on the under­ support the trend toward substituting communications for intra­ lying reason for air travel. Discussions with current users of tele­ company meetings. Several companies reported they have written conferencing suggest most of the existing applications are for intra­ policies restricting air travel, especially for meetings that only company meetings or for meetings involving companies that are involve other employees. Not all intracompany travel can be sub­ working together on specific projects. Sales calls still generally are stituted, for much of it requires physical presence, such as repair or held face-to-face, and communications technologies have yet to installation of equipment of programs. provide an alternative for large-scale conferences, conventions, and The possible small substitution of communications for business trade shows. air travel (3) and the forecasted 5 percent substitution by 2000 are About half of business travel is for meetings, convent~ons, and consistent with the regression equations presented previously in this trade shows (10). The remainder of business travel is fairly evenly report. They are also consistent with our understanding of the size . divided between intracompany business and intercompany busi­ and role of intracompany travel as a percentage of the total. Whether ness. The most likely candidate for replacement of business air or not communications can reach 25 percent or more substitution travel is intracompany business, or about 25 percent of air travel. will depend on technologies, customer acceptance, and costs, all of Discussions with travel representatives in several U.S. corporations which are still major unknowns. Wholesale trade

Air transportationt

Management and consulting services

State and local gov't, total

Federal gov't defense

Security and commodity brokers Business and professional associations

Legal services

U.S. Postal Service

Retail trade, except food service Colleges, universities, and professional schools

Federal gov't nondefense

Banking

Aircraft

Doctors and dentists D Communications

Freight forwarders and other •Air Transport transportation services

Periodicals

Newspapers

New and maintenance construction Real estate agents and services

Electronic computers

Electric services (utilities)

Engineering, architectural, and surveying services

Other business services

Automotive rental and leasing

0 1,CXXJ 2,000 3,000 4,000 5,000 6,000 Millions of Dollars

t Air Transportation includes imputed value of services provided by one airline to another. Source: U.S. Department of Commerce, Bureau of Economic Analysis, 1987 Input Output Tables.

FIGURE 3 Passenger air transportation and communications expenditures for selected industries, 1987. tAir transportation includes imputed value of services provided by one airline to another. Source: U.S. Department of Commerce, Bureau of Economic Analysis, 1987 Input Output Tables, 1987 Raphael and Starry 7

CONCLUSIONS The net affect of changing industry structure, cost control, and telecommunications on business air travel is uncertain. The indus­ There has been a definite downward shift in business air travel, as try shift, away from manufacturing and toward services should help substantiated by econometric analysis. The downward shift started increase enplanements, because of the faster rate of growth in sales in 1988, at a time when the economy started into a recession and and employment displayed by the service sectors. However, travel businesses began corporate restructuring and downsizing. Recent management and telecommunications are likely to place downward data confirm that the 1993 recovery did not produce the rebound pressure on business travel growth with the result that overall busi­ that would be expected with previous demand elasticities. ness enplanements and expenditures on air travel will not grow as We identified about 40 industrial sectors that account for 80 per­ fast as in the 1980s. cent of business air travel expenditures. We examined these busi­ nesses for several factors, including growth prospects and willing­ REFERENCES ness to use telecommunications. The results of this analysis support the contention that business air travel will not be as strong in the 1. Greens let, E. The Aviation Environment. Presented at 17th Annual 1990s and in previous decades. FAA Forecast Conference Proceedings, 1992. 2. Arvai, E. S. Telecommunications and Business Travel: The Revolution • Many of the traditional business travel industries are not per­ Has Begun. In Transportation Research Circular 425. TRB, National Research Council, Washington, D.C., May 1994. forming well or are forecast to grow slower than the national aver­ 3. Air Transport Association/Gallup Survey, 1994. age. Among these industries are defense, aerospace, and the U.S. 4. U.S. Department of Transportation, Federal Aviation Administration, Post Office. FAA Aviation Forecasts, various issues. • Most of the important business travel sectors spend more on 5. U.S. Department of Commerce, Bureau of Economic Analysis, The telecommunications than air travel. Through anecdotal evidence, it 1987 Benchmark Input-Output Accounts for the United States, Decem­ ber 1994, and earlier Input-Output Tables for 1982 and 1977. is clear many of the larger companies are frequent users of new 6. U.S. Department of Labor, Bureau of Labor Statistics. Outlook communications technologies, at least for intracompany meetings. 1990-2005, May 1992. Indeed, several companies have already mandated the use of 7. Chow, G. Tests of Equality between Sets of Coefficients in Two Linear telecommunications whenever reasonable or feasible. Regressions. Econometrica, Vol. 28, 1960, pp. 192-207. 8. New York Times. Expense Account Heyday Over, Companies Demand • Changes that are slowing the growth of air travel include Cheap Fares. July 27, 1994. downsizing of middle management; increased control over corpo­ 9. The American Express Survey, based on a mail questionnaire that was rate travel; and the spread of telecommunications, electronic mail, sent to U.S. businesses and government and educational organizations. and electronic data interchange (EDI) for communications and con­ The survey was fielded in March and April 1994. tact with suppliers, customers, and employees. In many industries, 10. Wada, I. Let's Make a Deal. Travel Weekly, August 18, 1994. the use of multiple communications technologies is changing how Publication of this report sponsored by Committee on Aviation Economics business people interact with each other. and Forecasting. 8 TRANSPORTATION RESEARCH RECORD 1506

Multiple Airport Systems in the United States: Current Status and Future Prospects

MARK HANSEN AND TARA WEIDNER

This study examines existing multiple airport systems (MASs) in the vices, consumers may benefit still further. Finally, airport competi­ United States and assesses the prospects for new MASs in the future. tion might allow less government interference in airport pricing and Using FAA and other data, we identify 14 MAS regions in the United investment decisions. Creager (2) argues that such airport dereg~la­ States, which account for 2.8 percent of all communities with commer­ tion is the natural counterpart of airline deregulation. cial air service and 43 percent of total enplaned passengers. The 14 MASs divide into roughly five clusters, based on their market size, and Despite these advantages, MAS development is opposed by a the concentration of traffic within them. Twelve of the MAS regions are number of factors. Travelers value service frequency, which is obvi­ "hubs" as defined by the FAA, whereas the others are MASs because of ously maximized when all flights leave from a single airport. There exceptional circumstances. Analyzing the 11 "large" and "medium" are also certain fixed station costs that encourage airlines to serve hub MASs, we model concentration, as measured by the Herfindahl only one airport in a region. Indivisibilities associated with other air­ concentration index, and find that it decreases with regional origin and port assets-runways, control towers, and so forth-may exert a sim­ destination (O&D) traffic and increases with connecting traffic. Next we use a binary logit model to find determinants of MAS status. We find ilar influence. A third consideration favoring a SAS is the economies that the probability of a region being served by a MAS increases with of hubbing, which depend on being able to fly passengers in and out the total traffic in the region, with some evidence that the probability of a single hub airport. Finally, there are political and institutional decreases with the ratio of total enplanements to O&D traffic. Using the barriers to the transition from a SAS to a MAS. These include resis­ two models and FAA forecasts for the year 2000, we find that 13 regions tance from incumbent airlines, political opposition to building a new currently served by a single airport have a significant MAS probability, airport, and risk averseness on the part of airport financiers. and that about half of these are likely to be fairly unconcentrated MASs. This report argues that, the above factors notwithstanding, con­ We conclude that the U.S. air transport system has reached the point in which MASs could become increasingly common, and in which airports ditions are ripe for a period of MAS development in the United could, therefore, become a competitive industry in many regions. States. We begin by inventorying existing MASs in the continental United States (Section 1). Next, in Section 2, we analyze the con­ sentration of passenger traffic in MASs, an indication of whether Many of the world's larger urban areas are served by more than one they are truly competitive. In Section 3, we consider the factors that commercial airport. Such multiple airport systems (MASs) offer determine whether an urban region is served by a MAS or a SAS. several advantages over single airport systems (SASs). Access costs Then, in Section 4, we assess the potential of existing SASs to are reduced, since travelers can choose the airport closest to their become MASs. Section 5 offers conclusions. true origin or final destination. This is not only more convenient, but also reduces social costs of vehicle travel, such as congestion and emissions. In addition, airports in a MAS will, ceteris paribus, be 1. MULTIPLE AIRPORT SYSTEMS: smaller than a single airport serving the same region. This implies AN INVENTORY reduced walking times, less costly parking facilities, and a less for­ midable wayfinding challenge for airport users. In short, when a This section identifies urban regions served by a MAS (MAS region is served by a MAS, the cost and inconvenience of getting to regions) in the continental United States. The identification of these the airport, and of getting through the airport, are reduced. regions involves a two-step process. Initially we identify regions More generally, a MAS offers the possibility for increased con­ with more than one commercial airport in the classification struc­ sumer choice and competition in the supply of airport services. As ture defined by the FAA (3). However, this list omits certain MAS Garrison and Gifford (J) point out, airline deregulation has given regions, while including others that, while nominally served by a passengers more choice with regard to airlines and service classes, MAS, have one airport that is so dominant that for all intents and but commercial airports continue to operate as state-run monopolies purposes it is a SAS. Therefore, we adjust the FAA list by consoli­ in most areas. The lack of airport choice, combined with the high dating regions into MASs, and redesignating certain highly con­ level of standardization among the products of most airlines, sharply centrated MASs as SASs. limit opportunities for air travelers to make choices and thereby Table 1 shows the 13 MASs recognized by the FAA in 1991 (3). reveal their service preferences. The mere existence of alternative This list includes all urban regions (termed "Communities" by the airports ensures somewhat greater consumer choice. If the airports FAA) served by two or more airports with scheduled commercial actively compete to provide the most attractive amenities and ser- air passenger service. We exclude cases in which the secondary air­ ports provide only air cargo, nonscheduled passenger, or general Air Transportation Research Center, Institute of Transportation Studies, aviation service, as well as regions located outside the continental University of California, Berkeley, Calif. 94720. United States. Hansen and Weidner 9

TABLE 1 MAS Airport Set Development Enplanement FAA CMSA/MSA Concentration Region Airport Set Additional Airports Index MAS Chicago, IL ORD, MOW, CGX 0.818 x New York City,, NY LGA, JFK EWR, ISP, SWF, HPN 0.313 x T Av C'l\.T A or TD T r.n Los Angeles, CA L~'\., 0J. ... rl., LIVA'-, 1.....1'-J.&J ONT, PSP, OXR, PMD 0.499 x Dallas/Ft. Worth, TX DFW,DAL 0.804 x San Francisco, CA SFO, OAK, CCR SIC, STS 0.529 x Washington, DC DCA, IAD BWI (1) 0.346 x Miami, FL MIA, FLL 0.608 x Houston, TX IAH, HOU, EFD 0.561 x Detroit, MI DTW,DET 0.937 x Santa Barbara, CA SBA, SMX 0.929 x Oshkosh/Appleton, WI ATW,OSH 0.830 x Seattle, WA SEA, BFI 1.000 (2) Tampa, FL TPA, PIE 0.999 (2) Philadelphia, PA PHL TTN 0.994 (2) Cleveland, OH CLE CAK 0.907 x Norfolk, VA ORF PHF 0.911 x Pensacola, FL PNS VPS 0.717 x

(l) Within 48 km (30 miles) of DCA in Washington DC MSA, an FAA '"large hub" community. (2) Not considered MAS due to extremely high enplanement concentration. Note: Index refers to Herfmdahl index of airport concentration, the sum of t11e squared airport market shares of t11e given passenger activity type. Sources: FAA Airports - FAA Airport Activity Statistics, 1991. CMSA/MSA Boundary - U.S. Bureau of the Census, 1990.

The set of MAS regions defined by the FAA is not complete. Sev­ among N airports, the HCI is 1/N. Within our set of MASs, the HCI eral major metropolitan areas are divided into multiple communi­ ranges from 0.9998, for Seattle, where one airport is highly domi­ ties. To rectify this, the FAA definition was modified to allow con­ nant, to 0.313 for New York. Three of the MASs-Seattle, Philadel­ solidation of airports located within the same 1990 Metropolitan phia (HCI = 0.999), and Tampa (HCI = 0.994)-have concentra­ Statistical Areas (MSA) or Consolidated MSA (CMSA) (4). This tions well above that for any of the others (the next highest is Santa action allowed for the consolidation of MASs in major metropoli­ Barbara, Calif., with an HCI of 0.929). These three MASs were tan areas as well as introducing four new MAS regions. Table 1 therefore redesignated SASs for purposes of our analysis. identifies these changes. We summarize the above discussion by offering a proposed def­ Additionally, an analysis was made of airports near all FAA inition of a MAS. In general we define a MAS as two or more air­ "large hub" communities. Baltimore-Washington International ports operating in a contiguous metropolitan area in such a way as (BWI) Airport emerged as a special case based on its close prox­ to form an integrated airport system. This integration is largely evi­ imity to the Washington, D.C. MSA airports (48 km, or 30 mi, from dent in the airports' competition for local passengers. More pre­ National Airport) and large traffic relative to these airports (27 per­ cisely, we define a MAS as consisting of two or more airports with cent of the combined total). Furthermore, these airports have been scheduled passenger enplanements, and which satisfy both of the planned cooperatively for many years, and the Baltimore and Wash­ following criteria: ington MSAs were in fact combined into a CMSA in 1991 (4). Con­ sequently, as shown in Table 1, we consolidate BWI with National . • Each airport is included in the same community by the FAA(3) and Dulles Airports to form a MAS serving the Baltimore­ or within 50 km (30 mi) of the primary airport of an FAA­ Washington region. designated "large hub" community, or each airport is in the same A final review reveals that our set of MASs includes several in MSA or CMSA (4); which a second airport handles a tiny amount of traffic relative to • The Herfindahl concentration index for the airports is less than the primary airport. Because of this large gap in activity, these com­ 0.95. munities effectively function as single airport systems. The Herfin­ dahl concentration index (HCI) was used as a measure of the degree Table 2 identifies the 14 MASs within the continental United to which passenger activity is concentrated at a single airport within States, based on the above definition. Enplanement and O&D pas­ the region. It is calculated as the sum of the squared traffic shares senger statistics for the MASs and their constituent airports are also of each airport in the MAS. ("Traffic" can mean enplaned passen­ presented. Most MASs have high traffic levels, with the majority gers or origin and destination (O&D) passengers. In identifying exceeding 10 million enplanements per year. However, there are MASs with excessive concentrations, we use the former.) For a also some small MASs, three with traffic levels under 0.5 million. SAS, the HCI is 1.0. For a MAS where traffic is evenly divided The MAS consists of two airports in eight cases, three airports in 10 TRANSPORTATION RESEARCH RECORD 1506

TABLE2 Existing Multiple Airport Systems

1991 Scheduled 1991 Domestic O&D Enelanements Enelanements FAA Airport Region and MAS Airports Hub Code Total Index Total Index CHICAGO, IL L 29,040,932 0.818 12,644,170 0.783 O'Hare International ORD 26,098,065 11,078,080 Chicago Midway MDW 2,935,166 1,564,190 Meigs Field CGX 7,701 1,900 2 NEW YORK CI1Y, NY L 27,918,360 0.313 19,806,200 0.329 Newark EWR 9,645,295 7,197,470 La Guardia LGA 9,121,466 7,998,160 John F. Kennedy International JFK 8,207,264 3,601,360 Islip/Macarthur ISP 410,150 452,000 Newburgh SWF 355,822 350,340 White Plains HPN 178,363 206,870 3 LOS ANGELES, CA L 26,276,420 0.499 20,113,310 0.321- Los Angeles International LAX 18,069,981 12,101,410 Ontario/San Bemadino/Ri verside ONT 2,831,551 2,729,010 Orange County/John Wayne SNA 2,544,596 2,450,970 Hollywood-Burbank BUR 1,821,400 1,803,720 Long Beach LGB 648,541 636,130 Indio/Palm Springs PSP 330,741 334,150 Oxnard/Ventura OXR 20,643 39,390 Palmdale/Lancaster PMD 8,967 18,530 4 DALLAS/FT. WORTH, TX L 25,416,828 0.804 9,286,950 0.635 Dallas/Ft. Worlh International DFW 22,625,338 7,052,420 Love Field DAL 2,791,490 2,234,530 5 SAN FRANCISCO, CA L 20,149,914 0.529 14,464,640 0.465 San Francisco International SFO 14,007,424 9,130,230 San Jose Municipal SJC 3,148,622 2,404,100 Metropolitan Oakland OAK 2,953,058 2,853,090 Santa Rosa/Sonoma County STS 35,675 65,690 Concord/Buchanan Field CCR 5,135 11,530 6 WASHINGTON, D.C. L 15,548,897 0.346 10,643,900 0.394 Washington National DCA 6,602,686 5,692,600 Dulles International IAD 4,706,395 2,452,140 Baltimore, MD BWI 4,239,816 2,499,160

three cases, and five, six, and eight airports in the remaining three The 14 MAS regions are diverse. In order to understand the dif­ cases. Enplanement distributions in the MASs are fairly concen­ ferent types of MASs that have developed in the United States, these trated, with half having HCI values over 0.8, and only two with systems were classified. The classification is based on Figure 1, in HCis under 0.5-the value for a two-airport MAS with evenly which the HCI and traffic of each MAS is pl.otted. For comparison divided traffic. The concentration is also evident in the fact that in "large hub" and "medium hub" SASs, all with HCis of 1.0, are also eight of the MASs, the busiest airport has more than five times as plotted. From this figure, five clusters of MASs are identified. These many enplanements as the second busiest. In no case, moreover, are characterized in Table 4. does a MAS include more than four airports with an enplanement Cluster 1, consisting of New York and Washington, D.C., is level more than 10 percent that of the busiest airport in the MAS. unique in its lack of a dominant airport. Enplanements are shared The MAS regions are compared to the FAA "hub" communities almost equally among three or more airports, as indicated by its low and the U.S. system as a whole in Table 3. The 14 MAS communi­ airport concentration indices ranging from 0.31 to 0.39. In contrast ties account for 43 percent of all U.S. passenger activity, even to the other MAS regions, these two exhibit more concentration in though they represent less than 3 percent of the communities receiv­ their O&D traffic than their total enplanements. This reflects the ing scheduled air service. All but two of the MASs lie in FAA­ fact that hubbing activity has focused on airports that are somewhat defined "hub" communities-those with at least 0.5 percent of total less competitive for the local market-Newark in the case of New U.S. enplanements. Indeed, 32 percent of the "large hub" commu­ York and Dulles and BWI in the case of Washington. Congestion at nities, which together account for 56 percent of "large hub" one or more airports in the MAS limits the market share of any one enplanements, are served by a MAS. In addition, 7 percent of the facility. "medium hub" communities and 2 percent of the "small hub" com­ Cluster 2 consists of five MAS regions, located in the South and munities have MASs. West. These differ from the first group in that in each case one air- Hansen and Weidner 11

TABLE2 Existing Multiple Airport Systems (continuetl)

1991 Scheduled 1991 Domestic O&D En12lanements En12lanements FAA Airport Region and MAS Airports Hub Code Total Index Total Index 7 MIAMI, FL L 12,575,026 0.608 7,503,320 0.526 Miami intemationai MiA 9,2i2,5i 7 4,609,900 Ft. Lauderdale-Hollywood Int'l H.L 3,362,509 2,893,420 8 HOUSTON, TX L 11,567,113 0.561 6,736,070 0.496 Houston Intercontinental IAH 7,805,317 3,428,090 W. Hobby HOU 3,756,251 3,275,790 Ellington Field EFD 5,545 32,190 9 DETROIT, MI L 9,791,172 0.937 5,121,710 0.883 Wayne County DTW 9,470,549 4,801,450 Detroit City DET 320,623 320,260 10 CLEVELAND, OH M 3,689,358 0.907 2,747,400 0.837 Hopkins International CLE 3,508,196 2,502,300 Akron/Canton CAK 181,162 245,100 11 NORFOLK, VA M 1,222,110 0.911 1,196,100 0.837 Norfolk International ORF 1,165,224 1,088,680 Newport News/Patrick Henry lnt'I PHF 56,886 107,420 12 PENSACOLA, FL s 450,473 0.717 507,280 0.618 Pensacola Regional PNS 373,688 376,770 Ft. Walton Beach VPS 76,785 130,510 13 SANTA BARBARA, CA 207,892 0.929 263,890 0.798 Santa Barbara SBA 200,269 233,820 Santa Maria Public SMX 7,623 30,070 14 OSHKOSH/APPLETON, WI 135,199 0.830 173,720 0.820 Outagamie County/Appleton ATW 122,515 156,310 Wittman Field/Oshkosh OSH 12,684 17,410

Note: Index refers to Herfindahl index of airport concentration. the sum of the squared airport market shares of the given passenger type. Sources: Enplanements - T3ffl00 Report, ONBOARD Database, ODProducts Inc., 1991. O&D- USOOT 10% Sample. Outbound Passengers, ODPLUS Database. ODProducts Inc .. 1991. port clearly dominates, resulting in a higher HCI, ranging from 0.50 by the traffic peaking due to connecting complex operations. to 0.61. However, secondary airports play a significant role in these The fourth MAS cluster features a lower total enplanement level MASs since the concentration indices are low compared to Clusters than the first three. The MASs in this group are also considerably 3, 4, and 5. Indeed, the· O&D traffic HCI for Los Angeles is very more concentrated, with enplanement HCis ranging from 0.91 to low, indicating the inability of any one airport to dominate the local 0.94. All of these MASs consist of only two airports, with the dom­ market in such a far-flung urban landscape. Houston, although con­ inant airport handling over 95 percent of the total passenger activ­ siderably smaller, is subject to much the same phenomenon. ity. This domination may be the result of the lower demand levels, Although this group also contains higher density areas such as San higher densities, or less severe congestion problems. Francisco and Miami, both of these have geographical features that The fifth and final cluster consists of MAS regions with very low encourage the dispersal of airport activity. In the case of the former, traffic. These MASs are somewhat less concentrated than the fourth San Francisco Bay results in longer travel distances than the popu­ group, with enplanement HCis ranging from 0.62 to 0.93. Two of lation density suggests, whereas the Miami region's location on a the three result from unusual circumstances. Pensacola is dominated narrow coastal plain, and resulting elongated form, has a similar by military activity which has encouraged ancillary passenger ser­ effect. As in the first cluster, airport capacity limitations may also vice to develop at a second airport. Likewise, Appleton is the site of reduce the concentration of enplanements. an airfield used by experimental aircraft. The Santa Barbara/Santa The MAS regions in the third cluster, Dallas and Chicago, are Maria MAS appears to be the result of these two formerly distinct distinguished by their role as major transfer points as the result of regions growing together. hub-and-spoke operations by air carriers. Because of the natural ten­ dency for transfer activity to concentrate at a single airport, the enplanement HCI is higher: 0.64 to 0.82. The large number of trans­ 2. DETERMINANTS OF MAS CONCENTRATION fer passengers at the primary airport results in a high service fre­ quency which serves to attract most of the local traffic. However, In this and the next section, we turn to statistical analysis of MASs the secondary airport in these regions is well located and attracts a in the United States. In this section, we analyze MAS concentration reasonable share of the local passenger demand, decreasing the levels, whereas in Section 3 we investigate the determinants of O&D HCI. These MASs are also subject to congestion, exacerbated MAS status. 12 TRANSPORTATION RESEARCH RECORD 1506

TABLE3 Comparative Activity Statistics

Number of 1991 EnQlanements 1991 MASs as Percenta8e of: Communities Total Domestic International Domestic O&D Large Hub Communities 32.14% 56.39% 55.27% 75.23% 59.31%. Medium Hub Communities 7.41% 7.85% 7.91% 4.24% 7.93% Small Hub Communities 1.82% 1.50% 1.54% 0.00% 2.03% Non-Hub Communities 0.52% 2.06% 2.06% 0.00% 2.39%

Total Hub Airports 12.73% 45.00% 43.83% 68.36% 44.10% Total All Airports 2.83% 43.24% 42.04% 68.26% 41.18%

Nok The definition of Multiple Airport System (MAS) communities consolidates several FAA "hub" communities.

We have already given considerable attention to the MAS concen­ tors will include anything that affects the distribution of traffic among tration, as measured by the HCI. The concentration is important for MAS airports. These include a host of airport attributes, such as loca­ several reasons. Since a SAS, by definition, has an HCI of 1, the HCI tion and accessibility, capacity, use restrictions, traveler and travel of a MAS measures how different it is from a SAS. The advantages agent awareness, and so on. These factors influence travelers' airport and disadvantages of a MAS as compared with a SAS become more choices directly, and also indirectly through their effect on airline ser­ significant as the MAS concentration decreases. In a highly concen­ vice supply decisions (5). Thus a complete representation of the trated MAS, passengers from throughout the MAS region will use the processes determining MAS concentration would require a highly primary airport, resulting in ground access costs comparable to a SAS. detailed analysis. There has been considerable work at this level (6-9), Likewise the primary airport of a highly concentrated MAS will be and more is ongoing, but here we want to paint with a broader brush. virtually as large as it would be in the case of a SAS. Also, for many At the macro level, we hypothesize that the level of concentra­ destinations, the primary airport of a concentrated MAS is likely to be tion in a MAS is affected by three factors. First, as local (or O&D) the only realistic alternative, depriving many or most air travelers of traffic increases, we expect concentration to decrease. More local the choice that we have argued is one of the main benefits of a MAS. traffic will result in more markets in which nonstop service can be Finally, because a highly concentrated MAS is not very competitive, supported from more than one airport. Since service quality gains for most policy purposes it must be regulated as a monopoly. The from flight frequency are diminishing, high local demand results in potential gains from airport deregulation are therefore lost. a reduced frequency advantage for the primary airport. Also, the Having established that MAS concentration is important, we con­ stronger the local traffic base, the more willing airlines will be to sider what factors may affect this variable. At a micro level, these fac- serve more than one airport in the region. Finally, the larger the pri-

1.000 • Santa Barbara ii Detroit 0.900 - : - • _Norfolk.. - • .Cleveland-

~ Chicago Q,j 0.800 "'O ·-·------.s Dallas/Ft.Worth c 0.700 0 l i Miami .... 0.600 -• - - ""c • San Fransisco Q,j • ~ 0.500 t ------c ... Los Angeles 0 u 0.400 'JJ • Washington, D.C. ~ ~ 0.300 ------l •-New York City I s0 0.200 Emo 0.100 -,- 0.000 l I I I I! I I I I I I I I I I 100,000 1,000,000 10,000,000 100,000,000 MAS Total Enplanements

l •Existing MASs •Other FAA Large and Medium "Hubs" I

Note: Index refers to Herfindahl index of airport concentration, the sum of the squared airport market share values of the given passenger activity type.

FIGURE 1 Total enplanements vs airport concentration Hansen and Weidner 13

TABLE4 Multiple Airport System Community Characteristics

Passenger Activi~ {millions) Land Area Ai!:Eort CaEaci!}:'.

Number of 1991 EPS 1991 Dom. O&D 1984 FAA Delays (I) AirEorts Total Index Total Index Area {~km~ 1990 2000 CLUSTER 1 New York City, NY 6 27.918 0.313 19.806 0.329 19,834 x x Washington, D.C. 3 15.549 0.346 10.644 0.394 17,032 x x

CLUSTER 2 Los Angeles, CA 8 26.276 0.499 20.113 0.327 88,081 x x San Francisco, CA 5 20.150 0.529 14.465 0.465 19,174 x x Miami, FL 2 12.575 0.608 7.503 0.526 8,200 x x Houston, TX 3 11.567 0.561 6.736 0.496 18,521 x x

CLUSTER 3 Chicago, IL 3 29.041 0.818 12.644 0.783 14,659 x x Dallas/Ft. Worth, TX 2 25.417 0.804 9.287 0.635 18, 122 x x

CLUSTER 4 Detroit, MI 2 9.791 0.937 5.122 0.883 13,481 x x Cleveland, OH 2 3.689 0.907 2.747 0.837 7,563 x Norfolk, VA 2 1.222 0.911 1.196 0.837 4,447

CLUSTER 5 Pensacola, FL 2 0.450 0.717 0.507 0.618 4,362 Santa Barbara, CA 2 0.208 0.929 0.264 0.798 7, 117 Oshkosh/Appleton, WI 2 0.135 0.830 0.174 0.820 3,670

Sources: Passenger Activity- ODPlus (T3ff100) and ONBOARD (10% Ticket Survey) databases, ODProducts, Inc., 1991. Index refers to Herfindahl index of airport concentration, the sum of the squared airport market share values of the given passenger activity type. Land Area - U.S. Bureau of the Census for CMSA/MSA. Delays - MAS contains airports to exceed 20,000 hours of delay in given year, Aviation System Capacity Plan, FAA, 199111992. mary airport in the region gets, the more it is subject to disec­ each airport will have a larger "captive" market area for whom it is onomies of scale such as longer walking distances or the need to by far the closest alternative. This will attract both passengers seek­ build expensive parking structures. ing convenience and airlines seeking a protected market niche. Second, as connecting traffic increases, we expect concentration To test these hypotheses; we estimated models of MAS concen­ to increase. An airline with a connecting hub in a MAS will tration. The models are of the form: naturally consolidate this operation at one airport, since it is prohibitively costly and inconvenient to transport connecting HCI ) - passengers between airports. For similar reasons, interline con­ In ( = O'. + f3 · ln(ODPAX) + 5 · ln(ENP) necting traffic is likely to concentrate at a single airport. If more I - HCI + A.· ln(AREA) + s (1) than one airline operates a hub in a MAS, the situation is somewhat more complicated, since the advantages of being able to offer inter­ where line connections by using the same airport must be weighed against HCI = Herfindahl concentration index; the traffic capture and product differentiation benefits from ODPAX =total MAS O&D passengers in the year 1991 operating hubs at different airports. In many cases, however, the (millions); latter is not an option, since only one airport in the region has the ENP = total MAS enplaned passengers in the year capacity to handle the traffic surges resulting from connecting 1991 (millions); complexes. AREA = land area of the region, in square miles; Third, as the land area of the MAS region increases, we expect con­ s = a stochastic error term; centration to decrease. When the land area is greater, ceteris paribus, O'., [3, o, and A. = coefficients to be estimated. 14 TRANSPORTATION RESEARCH RECORD 1506

The transformation of HCI in the above equation is used to convert tions) defined as "large" or "medium hubs" by the FAA (that is, a 0-1 variable to one whose range is -oo to +oo. According to our with enplanements of 0.25 percent or more of the national total), all hypotheses, J3 is negative, 8 is positive (since an increase in ENP, those with 20 million enplanements and 50 percent of the "hubs" controlling for PAX, implies an increase in connecting traffic), and with 10-20 million enplanements are MAS regions, whereas 90 A. is negative. percent with under 10 million enplanements are served by single The model was estimated on the 11 MASs included in the first airports. four clusters described above. The three MASs in the fifth cluster In addition to total traffic, it is reasonable to suppose that other were excluded, since they are unusual cases whose behavior is factors that affect MAS concentration also affect MAS status. unlikely to be captured by the same structural equation as the oth­ Indeed, we have already argued that a SAS can be viewed as an ers. Also, data for these small MASs are suspect since reported extremely concentrated MAS. Therefore, in light of the findings O&D traffic exceeds reported enplanements, probably because the from the last section, we expect regions with a higher proportion of latter excludes commuter carriers. (We did estimate the model with O&D traffic to be more likely candidates for MAS status. all 14 MASs, obtaining similar coefficient estimates and signifi­ On the other hand, there is a qualitative difference between a SAS cance levels, but with a slightly poorer fit.) and even a highly concentrated MAS. In the latter case, a second Estimation results appear in Table 5 which includes results for commercial airport exists, and in the first it does not. The barriers both the concentration of O&D traffic and enplanements. Both to this transition (in either direction: opening a commercial airport, models have fairly good fits, with adjusted R2 over 0.6. In both or closing it) are considerably higher than those to changing the dis­ cases, the land area variable is statistically insignificant and of the tribution of activity among existing airports. Thus we expect that at wrong sign. This suggests that land area is a poor measure of the any given time, the MAS status of a region will depend on past his­ phenomenon it is intended to capture-the extent to which airports tory as much or more than current conditions. This will introduce in the MAS have geographically protected market niches. The other considerable noise into the analysis of a snapshot of the system at a variables have the expected signs. O&D traffic is statistically sig­ given time. nificant at the 5 percent level, whereas enplanements is significant Notwithstanding this problem, we used a binary choice model to at the 10 percent level (and, in three of the four models, almost sig­ analyze the MAS status of the 51 large and medium hub airports. nificant at the 5 percent level). The O&D traffic coefficient has a The form of the model is: larger magnitude than the enplanement coefficient. This is reason.: ea,x,, able since an increase in O&D traffic implies an increase in P; (MAS)= _l_+_e_a-,x-;, (2) enplanements, but according to the arguments above should nonetheless result in a net reduction in airport concentration. where Finally, it is notable that connecting traffic is positively associated with O&D traffic as well as enplanement concentration. This P;(MAS) = probability that region i is a MAS region; implies that connecting traffic, itself tending to concentrate for rea­ X;k = a vector of regional characteristics; sons stated previously, pulls O&D traffic along with it. ek = a vector of coefficients to be estimated.

For the X;k we tried various combinations of enplanements, O&D 3. DETERMINANTS OF MAS STATUS traffic, their ratios, and their logs. Estimation results for three of the better performing models are summarized in Table 6. The models Next we consider whether MAS status, like MAS concentration, have reasonable explanatory power, with p2 (with respect of con­ can be explained at a macro level. Given a region, can we predict stants) approaching 0.5, and correct predictions for just over half of whether it is served by a MAS or a SAS? We already have some evi­ the MAS outcomes, and 90 percent of all outcomes. The models dence on this question from Table 3, which shows a strong correla­ confirm the relationship between total traffic and MAS status tion between hub class and MAS status. Further evidence is already observed. They also lend some support to the hypothesis obtained from categorizing "large" and "medium hubs" by enplane­ that MAS status is related to the O&D/connecting composition of ment level. Of the 51 regions (taking into account our consolida- traffic, although this effect is of marginal statistical significance.

TABLE 5 Estimation Results, Herfindahl Concentration Index Models

Independent Dependent Variable: Dependent Variable: Variable O&D Traffic HCI (Transfarmed) Enplanement HCI (Transformed) Constant l.90 2.56 l.87 1.89 (3.5) (4.1) (3.2) (3. 7) In(Enplanements) 2.12 l.94 1.57 1.57 (2.4) (2.2) (2.0) (2.2) ln(O&D Traffic) -3.75 -3.21 -2.63 -2.62 (-3.2) (-3.2) (-2.6) (-3.2) ln(Land Area) 0.48 0.01 (0.9) (0.0) R 2 (Adjusted) .64 .65 .61 .66 Note: t-statistics in parentheses. Land Area is in units of sqkm Hansen and Weidner 15

TABLE 6 Estimation Results, MAS Status Model (Binary Logit)

Independent Variable Model 1 Model 2 Model 3 Constant · -3.66 -0.93 -8.04 (-4. l) (-0.5) (-2. l) Total Enplanements (Millions) 0.27 0.41 0.38 (3.2) (2.6) (2.6) O&D Traffic/Total Enplancmcnts 5.15 (1.3) Total Enplanements/O&D Traffic -2.41 (-1.5) p 2 (zero) 0.56 0.62 0.59 p2 (constants) 0.41 0.49 0.46 Log Likelihood -15.59 -13.60 -14.36 No. Correct MAS Predictions 6/11 7111 6/11 No. Correct SAS Predictions 39/40 40140 39/40 Total Correct Predictions 45/51 47/51 45/51 Note: t-statistics in parentheses.

The latter is not surprising if one considers the data. In 1990, the 4. POTENTIAL MULTIPLE AIRPORT SYSTEMS nation had five regions with over 10 million enplanements and an enplanement/O&D ratio over 2-Chicago, Dallas, Denver, Atlanta, The models developed in Sections 2 and 3 were used to investigate and St. Louis. Of these, the first two are MAS regions and the oth­ the potential for additional multiple airport systems development ers are served by a SAS. All the regions with less than I 0 million over the next two decades. To do this, we calculated the MAS prob­ enplanements in 1990 and a enplanement/O&D ratio over 2 have ability of each of the 40 "large" and "medium hubs" currently SASs but so do most other regions with traffic below this threshold. served by a single airport. We did this both for 1990 and, using the Thus, although the evidence that connecting traffic affects MAS FAA forecasts (JO) for the year 2000. For purposes of comparison, concentration is quite strong, it is far less definitive in the case of the probabilities were also calculated for existing MASs. MAS status. Figure 2 shows the I 990 and 2000 results for all existing MASs and those SASs with a nontrivial MAS probability. The 1990 results show that, among the existing MASs, Detroit, Cleveland, and Nor-

90% • • •0 • • • • • 80% 0 • • Potential MASS • • • • <> • 70% • • • • • ~ 60% 0 o 1990 Model 1 :c 0 • o 1990 Model 2 D • "'0 50% o 1990 Model 3 Q: • • • • 2000 Model 1 Cl) 0 • • ct 0 0 • • 2000 Model 2 :::: 40% <> • • <> 0 0 • • 2000 Model 3 0 0 <> • • • 30% 0 • <> 0 • • 0 0 0 0 <> <> 20% 0 Existing MASs <> 8 <> 0 a •<> 0 B 0 10% 0 0 8 0 •

fJ) 0 x fJ) 0 QJ c c fJ) QJ c .!!! QJ 0 Cl ·2 0 0 QJ -~ :c"' E Cl Qi c iii a. ~

FIGURE 2 1990 and 2000 range of logit model predictions 16 TRANSPORTATION RESEARCH RECORD 1506 folk are the least probable ones, based on their low traffic levels. craft delay in 1990. Consequently, additional capacity is likely to be These are all highly concentrated MASs. Both Cleveland and Nor­ needed. The predicted enplanement concentration gives an indica­ folk are also cases in which the airports operate under separate local tion of how much traffic might be diverted from the existing airport jurisdictions between which there is significant rivalry. This may if additional capacity were supplied in the form of a new airport help to explain why activity has not been consolidated at one airport (assuming the existing airport continued operating). Phoenix, in these regions. Among the SASs, Atlanta is the most likely can­ Boston, Orlando, Las Vegas, Philadelphia, and San Diego have the didate to be a MAS, based on 1990 data. Denver, Phoenix, Boston, lowest concentrations. In all of these cases, our model predicts a siz­ and Las Vegas also have a fairly high MAS probability, although it able (30 percent or more) diversion of enplanements from the exist­ is under 50 percent in all cases. ing airport under the MAS scenario. On the other hand, Atlanta, Figure 2 also shows MAS probability estimates for the year 2000, St. Louis, Minneapolis, Pittsburgh and Charlotte would be expected based on traffic growth projected by the FAA and assuming a ratio to have such a highly concentrated MAS that building an additional of enplanements to O&D traffic equal to that in 1990. The contrast airport would hardly be worthwhile. Denver and Seattle are inter­ with the 1990 results is striking. Six regions-Atlanta, Denver, mediate cases, in which MAS development would disperse traffic Phoenix, Boston, Las Vegas, and Orlando-are forecast to realize to some degree, although the primary airport would continue to be traffic growth that would give them a MAS probability of over 50 dominant. (Denver, of course, has elected to build a new airport but percent based on two of the three models. Two others-St. Louis close the existing one.) and Philadelphia-surpass the 50 percent threshold according to The predicted O&D traffic concentration measures the extent to one of the models, whereas three other regions approach it. These which passengers originating in each hypothetical MAS region results form the basis of the assertion made at the outset­ would have a real choice among airports, and also. the degree to conditions appear to be ripe for substantial increase in the number which having multiple airports would reduce the private and social of multiple airport systems in the United States. costs of ground access. It is consistently lower than the enplanement Figure 2 shows considerable variability in the MAS probabilities concentration, suggesting that a MAS can yield sizable consumer predicted by the different models. The major uncertainty behind this choice and ground access benefits even when the overall traffic variability is whether the distribution of traffic between connecting remains highly concentrated. and O&D passengers affects the probability of being a MAS. Regions with unusually high or low proportions of connecting traffic-St. Louis and Boston, for example-thus show the highest 5. CONCLUSIONS variability. Predictably, Model I, which is based on enplanements only, yields the highest MAS probability for St. Louis, whereas the Our results suggest that the potential exists for multiple airport sys­ models that incorporate a connecting traffic effect yield the higher tems to play a significantly expanded role in the U.S. commercial MAS probabilities for Boston. airport system. We have identified 13 SAS regions whose traffic Table 7 provides further indications of the potential for MAS growth and traffic composition make them potential candidates for development in the SAS regions included in Figure 2. It identifies MAS development. Of these, roughly half could benefit substan­ those that are, or are forecast to be, subject to significant amounts tially from such development in terms of traffic diversion, consumer of aircraft delay, and, using the concentration models discussed in choice, and ground access cost. All of these regions have commer­ Section 2, predicts what the enplanement and O&D traffic concen­ cial airports that have or are expected to soon have significant air­ tration levels would be if a MAS developed. (In applying the con­ craft delay. Serious consideration should be given to expanding air­ centration models, we again assumed a ratio of enplanements to port capacity in these regions by adding a second airport. O&D traffic equal to that in 1990). With the exception of Las Deciding whether to build a second airport is bound to be Vegas, airports in all these regions had significant amounts of air- · controversial. Political conflict appears in the present era to be an

TABLE7 Potential MAS Regions

FAA Forecast Projected O&D Over 20,000 Range of Predicted MAS Predicted Concentration Enplanements, Traffic, Hours of Delay Statust Probabilities (HCI) if Region were MAS Region 2000 2000 1990 2000 Min 2000 Max 2000 EPS O&D Traffic Atlanta, GA 24.6 8.7 x x 90.7% 95.6% 0.86 0.78 Denver, CO 21.4 9.6 x 88.5% 92.1% 0.77 0.68 Phoenix, AZ 15.6 10.1 x x 62.4% 84.7% 0.62 0.54 Boston, MA 13.7 11.0 x x 49.6% 84.0% 0.48 0.43 Las Vegas, NV 14.1 10.4 x 52.8% 82.7% 0.55 0.48 Orlando, FL 13.9 9.9 x x 48.0% 78.3% 0.56 0.49 St. Louis, MO 14.6 5.9 x x 29.4% 56,0% 0.89 0.81 . Philadelphia, PA 10.4 7.7 x x 29.0% 51.2% 0.64 0.56 Minneapolis/St. Paul, MN 13.4 6.1 x x 31.6% 47.9% 0.86 0.78 San Diego, CA 8.3 7.6 x x 19.0% 46.0% 0.54 0.47 Seattle, WA 11.3 7.1 x x 34.2% 46.7% 0.72 0.63 Pittsburgh, PA 12.8 4.3 x x 5.8% 43.6% 0.94 0.89 Charlotte, NC 11.2 2.3 x x 0.0% 33.8% 0.99 0.97 Source: Forecast EPS - Terminal Area Forecasts, FAA, CY1993. Delays - Aviation System Capacity Plan, FAA, 1991/1992. Hansen and Weidner 17 inevitable concomitant of any major addition to airport capacity. ACKNOWLEDGMENT Savas (11) has observed that decisions involving urban systems frequently involve the search for a single politically feasible solu­ This research was sponsored by the California Department of tion instead of choosing the best from a set of fea~ible alternatives. Transportation, Division of Research. This clearly applies to airport planning. Our findings help establish the conditions under which a MAS may be viable, and thus when alternatives involving more than one airport can be added to the list of candidate solutions to be subjected to the political feasibility REFERENCES test. For example, it may be possible to build a more modest airport and keep the existing one operating instead of constructing a huge l. Garrison, W. L., and J. L. Gifford. Airports and the Air Transportation Systems: Functional Refinements and Functional Discovery. In Tech­ new airport and close the existing one, or avoid a major expansion nological Forecasting and Social Change, Vol. 43, 1993, pp. 103-123. of an existing airport in favor of adding capacity by means of a new 2. Creager, S. Airline Deregulation and Airport Regulation. In The Yale airport. By no means will these MAS alternatives be the preferred Law Journal, Vol. 93, 1983, pp. 319-339. ones in all situations, but in some cases they may allow capacity 3. Airport Activity Statistics of Certificated Route Air Carriers, CYJ991. expansion to go forward when otherwise it would be politically Series: FAA-AP0-92. Federal Aviation Administration, Research & Special Programs Administration, 1992. infeasible. 4. Statistical Abstract of the United States. U.S. Department of Congress, If the near term benefit of MAS development is capacity expan­ Bureau of the Census, 1990 and 1991. sion, in the long term there is also the potential of creating a com­ 5. Hansen, M., and Q. Du. Modeling Multiple Airport Systems: A Positive petitive, entrepreneurial airport sector. In addition to permitting the Feedback Approach. Research Report 93-12. Institute of Transportation Studies, University of California at Berkeley, 1993. obvious locational choices, this would encourage greater experi­ 6. Gelerman, W., and R. De Neufville. Planning for Satellite Airports. In mentation with services and prices, and thus increase the range of Transportation Engineering Journal, Vol. 3, 1973, pp. 537-551. air travel options available to consumers. Put another way, airport 7. Harvey, G. Airport Choice in a Multiple Airport Region. In Trans­ deregulation is desirable for many of the same reasons that airline portation Research, Vol. 21, No. 6, 1987, pp. 439-449. deregulation was. MASs, by curtailing the ability of airports to 8. Innes, J. D., and D. Doucet. Effects of Access Distance and Level of Service on Airport Choice. In Journal of Transportation Engineering, engage in monopolistic abuses, improve the prospects for such Vol. 116, 1993, pp. 507-516. deregulation. 9. Multiple Airport Demand Allocation Model (MADAM) Calibration and As Garrison and Gifford (I) point out, the air transportation sys­ Sensitivity Analysis. National Capital Region Transportation Board, tem is in many respects mature, with highly standardized products 1985. 10. Terminal Area Forecasts FY 1993-2005. Series: FAA-AP0-93-9, U.S. and relatively little innovation. As air transport demand increases, Department of Transportation, FAA, Office of Aviation Policy, July the tendency is to respond by growth instead of development, by 1993. adding capacity to SASs instead of creating MASs. It is important 11. Savas, E.S. Systems Analysis and Urban Policy. In Systems Analysis in to realize that this is not the only option, and that the MAS alterna­ Urban Policy-Making and Planning (M. Batty and B. Hutchinson, tive could ultimately lead to a more innovative and dynamic air eds.), Plenum Press, New York, 1983. transportation system. We see an opportunity, and a challenge, in Publication of this paper sponsored by Committee on Airfield and Airspace cultivating a competitive airport sector from monopolistic roots. Capacity and Delay. 18 TRANSPORTATION RESEARCH RECORD 1506

Forecasting Air Travel Arrivals: Model Development and Application at the Honolulu International Airport

SANJAY KAWAD AND PANOS 0. PREVEDOUROS

Honolulu International Airport is one of the dozen busiest airports in the taken from the United Nations (UN), the Organization for Economic United States. It provides a port of entry for foreign flights and both Cooperation and Development (OECD), the International Monetary domestic and interisland terminals. For Hawaii's airport engineers and Fund (IMF), the Economist Intelligence Unit (EIU), the Pacific Asia planners, this status makes short- to medium-term air travel forecasts Travel Association (PATA), and World Bank publications. Variables essential, particularly for landside applications. A unique model system is described with separate models for each origin region or country. The representing other factors affecting travel behavior (e.g., wars, models are estimated with the Cochrane-Orcutt regression procedure. exchange rates, and promotional fares) also were introduced. Major explanatory variables include the gross national product (GNP), Two independent concepts for model estimation were estab­ gross domestic product (GDP) (in various forms), and the consumer lished. The first entails the estimation of global air-travel generation price index (CPI). Exchange rates, strength of currencies, and variables and the estimation of Hawaii's (variable) share of that market. This for wars, recessions, and airline strikes are also introduced. The models process explicitly accounts for the effects of competition and is adhere closely to the actual number of arrivals, and all variables perform as expected. The USA model has an overall error rate (for the 18 years under consideration for future investigation. The second entails the of the estimation span) of less than 0.05 percent and annual extreme estimation of separate forecasting models for each major originat­ errors ranging from -6.4 to 5.7 percent. The Japan model also has an ing point of air travel to Hawaii. Competition is accounted for in the overall error rate of less than 0.05 percent and annual extreme errors number of arrivals from a particular country or region. Models for ranging from -7.5 to 10.4 percent. Near-future values for explanatory arrivals from the mainland United States and Japan are presented. variables can be obtained from major economic organizations, thus the Arrivals from these places of origin constitute approximately 75 use of the model system requires only that planners update variable val­ percent of the total number of arrivals to Hawaii. ues annually from regularly issued publications. Long-term forecasts are more difficult due to the lack of published data. In those cases, the This study is unique in its consideration of economic, demo­ user must resort to a combination of trend extrapolation with ARIMA, graphic, and other defining characteristics of the originating regions as shown in this discussion, and educated estimates based on contem­ and countries. As noted in the literature review, such a model sys­ porary macroeconomic literature. tem has not yet been developed. This study consists of five parts. After the introduction, a litera­ ture review focusing on forecasting methodologies and destination The development of an air travel demand model system for Hawaii competition is presented. Concept development and methodology is discussed. Hawaii's airport system includes five major and sev­ (including data description and variable definitions) are then dis­ eral secondary airports. It is owned by the state and managed by the cussed, followed by an explanation of the USA and Japan models Airports Division of the state department of transportation (DOT). with forecasts to the year 2000. The final section presents conclu­ The Honolulu International Airport is one of the busiest in the sions and directions for future research. nation; in 1992, it ranked ninth in the number of domestic passen­ gers and fourth in the number of international passenger arrivals (category X airports, American Airport Traffic Report 1993). LITERATURE REVIEW Because Hawaii's airport system is the primary mode of passen­ ger transportation into and out of the state, accurate demand fore­ The review of literature that follows focuses on (a) air travel fore­ casting is essential. Forecasts are used for landside planning appli­ casting methodologies (estimation procedures and model specifica­ cations (e.g., ramp control, baggage handling, pedestrian corridor tions) and (b) the major explanatory variables that have been used level-of-service, queueing analysis at check-in counters and secu­ in past studies and proven useful. The review also covers destina­ rity points, size of holding areas, parking demand, and traffic cir­ tion competition. Such a discussion is appropriate for destinations culation volume) and by the Airports Division as input to its Airport with a high number of tourist arrivals. [A more detailed presentation Landside Planning System (ALPS) software. of existing literature can be found in a thesis by the first author (J)]. The research goal was the estimation of user-friendly air travel forecasting models for the Airports Division. For ease of use and reli­ ability of forecasts, the models were estimated with macroscopic data Forecasting Literature

Although several studies to model tourism in Hawaii have been S. Kawad, Parsons Brinkerhoff Quade and Douglas, Inc., Two Waterfront Plaza, Suite 220, 500 Ala Moana Boulevard, Honolulu, Hawaii 96813. P. D. undertaken (2,3), none estimate demand from originating countries Prevedouros, Department of Civil Engineering, University of Hawaii at based on the economic and demographic variables of those coun­ Manoa, 2540 Dole Street, Holmes Hall 383, Honolulu, Hawaii 96822. tries or the competition between tourist-attracting countries. Kawad and Prevedouros 19

The Hawaii Visitors Bureau (HVB), a nonprofit organization that • The distance between generating country i and recipient coun­ promotes Hawaii as a tourist destination, forecasts visitor arrivals try j; and for the short-term (i.e., for a period of 1 year). HVB uses (a) survey • The value of time for n = 1963, 1967, 1975, 1980. data from a sample of travel agents worldwide, (b) immigration records, and (c) data from continuously conducted in-flight surveys. It is not clear how the value of time is defined. Data from the travel agents reflect committed (not future) travel Two base years ( 1963 and 1967) were considered and the model plans, and the immigration and in-flight survey data offer insights gave tourist flows for the years 1975 and 1980 between each of the concerning the actual situation of visitor arrivals. 522 pairs of generating and recipient countries. The forecasts were The Department of Business, Economic Development, and verified by conducting a time-series analysis of the actual number Tourism (DBEDT) uses a model (4) to project visitor arrivals based of tourists generated by each origin country as a function of its pop­ on assumptions about factors such as air fares and income levels, ulation, GNP per capita, and average distance traveled abroad. which have failed to account for the (largely worldwide) economic Armstrong concluded that the biggest increases over the forecast recession of the early 1990s and the "air-fare wars" among airlines period 1967 to 1980 would be in arrivals to nontraditional tourist in recent years. The DBEDT model considers the economies of the countries (namely, Fiji, New Zealand, Australia, etc.). U.S.A. and Japan markets only. Markets such as Korea, the Philip­ Witt and Martin (8) examined the choice of appropriate vari­ pines, Australia, New Zealand, Canada, and Europe are also impor­ ables to represent tourists' cost of living for various origin-and­ tant to Hawaii, but are not considered in the DBEDT model. The destination countries. The model was an attempt to test whether the main purpose of the DBEDT model is to forecast the state's econ­ consumer price index (CPI) is a suitable proxy for the cost of Ii ving omy and act as a source of information for land transportation mod­ of tourists at the destinations. Their model attempts to explain the els and highway planning. flow of tourists from major origin countries [namely, France (to Stuart (5) used the gravity model to interpret air-passenger traf­ Portugal and the United Kingdom), Germany (to Austria and Italy), fic between Hawaii and the mainland United States. The aim of this the United Kingdom (to Spain and Greece), and the United States study was to test the hypothesis that air-passenger traffic is a func­ (to Canada and France)]. A foglinear model structure was used. The tion of distance and population distribution. The model was not lost of the time-series explanatory variables includes the population used to predict the future pattern of passenger movement to Hawaii. and personal disposable income of each origin country; the cost of A review of the literature reveals that much more work on avia­ living for tourists at each destination; a weighted average of the cost tion forecasting has been done outside Hawaii. The objective of the of tourism in substitute destinations; the exchange rate of the cur­ model specified by Crouch et al. (6) was to estimate the impact of rencies between each origin-destination pair; air fares; a weighted international marketing activities of the Australian Tourist Com­ average of air fares to substitute destinations; the cost of travel by mission (ATC) on the number of tourist arrivals. Multivariate log­ surface; and dummy variables. linear regression analysis was used to estimate the elasticities of The authors concluded that tourist cost-of-living data were not demand by considering the economies of the United States, Japan, superior to the simple CPI or simple exchange-rate proxies. The New Zealand, the United Kingdom, and Germany. The model empirical results indicated that the CPI (either alone or with the includes: exchange rate) is a reasonable proxy for the cost of tourism (8). The literature shows that the GNP, CPI, and dummy variables • Real per-capita disposable personal income (which is used as representing airline strikes, recessions, and exchange rates are an approximate measure of the price of tourist services in Aus­ important determinants of intercity travel. tralia); • Air fares to Australia from the origin country; • Promotional expenditure by the ATC; Destination Competition • A trend term to allow for changing "tastes"; and • Dummy variables to account for the effect of special condi­ Tourism is one of the most rapidly growing sectors of the interna­ tions in certain years. tional economy. As more countries realize the importance of tourism to their economies, the competition among them will The model reflects the historical data well, but its forecasting increase. The U.S. market has a large share of the world tourism ability is questionable given the lack of knowledge for forecasting market and Hawaii's potential for growth in tourism is great. several of the independent variables. In a study of the destinations competing with Hawaii (9), places Armstrong (7) used cross-section analysis and a logarithmic with the same attributes (such as sun, sand, and sea) were chosen. model form to forecast international tourism demand. Model vari­ Additional criteria were established to narrow the list. The ables include the number of tourist arrivals recorded in country j Caribbean, Mexico, and Australia emerged as the primary destina­ coming from country i (dependent variable), a value that is assumed tions competing with Hawaii. to reflect the tourist appeal of country j and which arises from fac­ Studies also (10-12) have shown that natural beauty and climate tors not explicitly included in the model. Those factors include: are the primary elements that attract people to a destination. Addi­ tional factors include culture, sports, shopping, and night life. • Climate; Although the potential worth of attractiveness to modeling • Resorts and culture; arrivals is obvious, the incorporation of such factors in time-series • The population of country i; models is problematic for two reasons. First, definitions of attrac­ • The income per capita of country i [gross national product tiveness and historical measurements of it are lacking. Elements of (GNP) per capita]; attractiveness that remain constant over time (e.g., the cultural value • A value given to the proximity of frontiers or common lan­ of historical sights and major museums) are not useful in time­ guage, if any, between countries i and}; dependent models. Elements of attractiveness that change over time 20 TRANSPORTATION RESEARCH RECORD 1506

(e.g., crowding of beaches, congestion, and pollution) are useful in hence continuously "corrected" forecasts can be obtained for sensi- time-series models if (a) consistently measured over-time indices ble short- and medium-term planning applications). . (or reasonable proxy variables) are available, and (b) future values Finding accurate time-series data of arrivals from a given region for these indices can be forecast with reasonable confidence (i.e., or country is a major challenge because the arrivals often include easier to forecast than the dependent variable). There is a clear lack travelers who made a connection at the region. For example, most of research in this area, and to the best of the authors' knowledge, visitors to Hawaii from Hong Kong, Taiwan, Singapore, and China the aforementioned conditions are not met. arrive via Tokyo. Therefore, sources that apply the appropriate Second, the real issue may not be the actual quantitative assess­ screenings (from travel agent bookings and on-board surveys) were ment of a dimension of attractiveness (which could be estimated selected for the collection of arrival data. Specifically, arrival data with a level of service or quality index), but its perception in vari­ from 1974 to 1983 were taken from PATA reports (13), and data ous markets. Perceptions not only vary from reality, but also differ from 1983 to 1992 were obtained from HVB publications. among cultural or ethnic groups. (One might ponder the perception The generic structure of the models is given as of safety from crime in Los Angeles by Tokyo and New York City residents). Thus, perceptions add another layer of difficulty to an ARRIVALS~ = ARRIVALS~ - 1 • (I + %CHANGE~ - 1 to N) already difficult problem. %CHANGE;N _ 1 10 N= f (GDP or GNP, CPI, CURRENCY Because the goal of the research was to estimate a set of robust EXCHANGE, STRIKE, WAR, and user-friendly models for engineering and planning applications, AIR FARE, ... ) and because of the issues discussed in this section, indices of attrac­ tiveness were not used. where

ARRIVALS~= visitor arrivals in year N, for country METHODOLOGY or region i;

ARRIVALS~- 1 =visitor arrivals in year N - 1, for Model Structure country or region i;

%CHANGE~_, to N =model estimate of the percent change The initial modeling concept focused on the estimation of global from year N to N - 1, for country or air-travel generation, followed by the estimation of Hawaii's (vari­ region i; and able) share of the international market. This process explicitly f(GDP, ... ) = model specification (the list of accounts for the effects of competition. The estimation of a global explanatory variables in each model air-travel generation model is still at the conceptual level because is presented later). the collection of global tourism data is difficult and good definitions The following steps were part of the process of model develop­ of competing destinations are not available. Theoretically, all ment: tourist destinations compete for visitors. The authors do not agree that only destinations with sun, sand, and sea, as concluded in the 1. Data collection and selection of variables, study conducted for the DBEDT by Arthur Young, Inc. (9), com­ 2. Estimation of alternative model specifications and statistical pete with Hawaii. Many Americans may decide to visit Europe testing, and instead of a seaside resort when the dollar is strong and the air fares 3. Model refinement and selection. to Europe are similar or less expensive than those for Hawaii. Besides the Caribbean, Mexico, and Australia (9), Florida and California are formidable competitors (particularly for families with children), not only because of lower total air fares, but also Estimation Procedure because of the abundance of family-oriented theme parks in those states. For similar reasons, other destinations in the Pacific (e.g., The equations were initially estimated using the ordinary least Indonesia, the Polynesian islands, Singapore, and Thailand) should squares method. In all cases the Durbin-Watson (DW) statistic indi­ be included when considering Asian markets. Thus, the problem cated the presence of autocorrelation; thus, the parameter estimates of destination competition becomes complex and impossible to were inefficient and the regression assumptions were not valid. The address with the scarce, often incompatible, and short-span time­ models were reestimated using the Cochrane-Orcutt iterative pro­ series data available (n.b., properly adjusted and defined macro­ cedure to reduce the likelihood of autocorrelation. The following economic statistics are not available before the 1980s for several criteria were used to arrive at the final models: countries): The actual approach for model development involved a model 1. Correct signs for the coefficients: the GNP-GDP variable system with separate forecasting models for each major origin should have a positive sign, the CPI variable should have a negative country (or region) of air travel to Hawaii. Competition is partly sign, the exchange rate coefficient should be positive, and the accounted for in the dependent variable, which is the number of dummy variables should have logical signs. arrivals from the particular country or region. It cannot be pre­ 2. DW statistic: a DW statistic lying between 1.8 and 2.2 is used sumed, however, that preferences represented in the dependent vari­ as a measure for no autocorrelation. A DW statistic equal to 2 indi­ able will be perpetual. Hence, forecasts beyond a 5- to 10-year cates the absence of autocorrelation (8). period should consider destination competition explicitly, or the 3. t-statistic and R2 value: the statistical significance of parame­ models and their estimates should be updated regularly. (Specifi­ ter estimates and the ability of the model to explain a large portion cally, the models were designed for forecasts up to 10 years and of the variance of the dependent variable are important indicators of with the stipulation that the coefficients will be updated annually; goodness-of-fit. However, models with the highest R2 were not nee- Kawad and Prevedouros 21 essarily selected as best if other criteria were violated. Also, vari­ Certainly, air travel is a multifaceted (multivariate) phenomenon; ables with not-significant parameter estimates were retained when thus, a consistent correlation between arrivals and macroeconomic a reasonable justification was available. variables is not observed. As shown in the next section, however, the developed specifications (which include several of the afore­ mentioned variables) result in a model that explains more than 75 Explanatory Variables percent of the variance of the dependent variable. The list of variables used in the specifications of the USA model Several mostly macroeconomic variables were considered in the follows. model specifications. Figure 1 presents several variables pertaining to the USA model. The dollar index of relative strength to a num­ LLGDPC =Annual change of the gross domestic product per ber of foreign currencies reported by the Chicago and New York capita in percent; performed best in a logarithmic currency markets (monthly futures) shows that, at times when the transformation and lagged. Specifically: dollar is strong (e.g., 1984 and 1985), a decrease in arrivals is LLGDPC = 0.75 · Log(GDPCN _ 1) + 0.25 · Log(GDPCN _ 2) observed as foreign travel becomes more affordable. Conversely, Thus, the per capita product of up to 2 years before when the dollar index dropped sharply in 1986, a sharp increase in the target year affects arrivals. The parameters of arrivals to Hawaii was observed. 0.75 and 0.25 were obtained with a trial-and-error The annual change in arrivals and in the CPI show roughly oppo­ process with values from 0 to 1. The combination site trends, as expected. Low inflation promotes discretionary shown yielded the best model fit. expenditures such as vacations, whereas high inflation (e.g., 1978 CPI = Annual change of the consumer price index. to 1980) curtails them. STRIKE = Dummy variable to account for the United Air­ The annual change in arrivals and in the per-capita GDP show lines strike in 1985; this event caused a signif­ similar trends. A strong economy generates business and discre­ icant drop in visitors to Hawaii given that tionary travel. Economic declines (e.g., 1987 to 1990) and reces­ United's approximate market share in Hawaii is 30 sions (e.g., 1981and1990) have a strong detrimental effect, which percent. often lags by 1 year. For example, the poor economic performance NATURE = Dummy variable to account for the devastation in 1979 resulted in a decrease in arrivals in 1980. caused on Kauai by Hurricane Iniki.

Annual Change in Arrivals U.S. Dollar Index

16 150 12 8 4 i 0 (.I t -4 Q. -8 -12 -16

-20+---+----+--+--+--+---+---+__.~>---+---+--+--+---+--+--+---< '

Annual CPI Change Annual Per Capita GDP Change

140 12.0 I.

10.0 I c ~ 8.0 (.I

~ ""Q. 6.0 4.0 2.0 0.0 '

FIGURE 1 Trend of arrivals from the mainland United States and selected factors affecting them. 22 TRANSPORTATION RESEARCH RECORD 1506

DJ/A = Dummy variable based on the widely publicized 3. A dummy variable to account for years when the U.S. dollar Dow Jones Industrial Average index (DJIA) and is particularly strong (e.g., 1984 and 1985). used as a proxy of "economic mood." A positive 4. A variable based on Department of Commerce statistics on economic (and spending) mood is assumed (value year-over-year change of disposable incomes in the United States. of + 1) when the index gains 100 or more points 5. A variable based on the average annual price of a barrel of within a year (difference between the last and first crude oil. trading day for the year); a negative mood is 6. Values of air fares, both current and constant, were used with assumed (value of -1) when the index loses 100 or little success, partly because air fare increases (and decreases in the more points in a year; and a neutral mood is early 1990s) follow the fluctuations in demand. In addition, the assumed (value of 0) when the index gains or loses large proportion of frequent-flier seats used for travel to Hawaii fur­ fewer than 100 points in a year. The use of this (and ther distorts the picture of air travel cost. similar) indices may be partic~larly appropriate for high-cost destinations, which are afforded by Figure 2 presents two major variables included in the Japan mostly higher-income households, many of which model and the dependent variable. The arrivals from Japan and the are likely to be investors and, thus, will follow the GNP show similar trends; for example, in the years of GNP decline economy's progress represented by the DJIA. (1983 and 1986), arrivals also decline. The only inconsistency in the two trends is the span between 1975 and 1979, when GNP growth Variables tested but excluded from the final specification declines but arrivals increase (partly because of the yen's strength). included: As a result, a dummy variable (175-79) was introduced to account for the inconsistency, which can ultimately be explained by a lack 1. The annual change in the gross state product (GSP) of the state of vacation time and the unlikelihood of foreign travel, both of of California. It was observed that when the annual growth of Cal­ which have changed substantially since the early 1980s, at least in ifornia's GSP drops below 3 percent, a decrease in arrivals is Japan's urban centers. observed. California is a very important market for Hawaii, partly The yen index serves as a signal to growth in arrivals from Japan. because of its market size, proximity, least-expensive air fares, and When a large gain in the exchange rate is observed, arrivals frequent service. Ultimately, a separate model for the West Coast is increase. Conceivably, large increases in the value of the yen vis-a­ sought, but the procurement of arrival data that do not include vis the dollar "make the news" in the Japanese market, which stopovers in that region so far has not been possible. responds positively when an attractive product (a vacation or wed­ 2. The Persian Gulf war of 1991. ding in Hawaii) becomes more affordable in yen.

Annual Per Capita GNP Change Yen Index

11 0.9 c Q,I /'. ;.... Q c ~ 0.7 ,·"--.·· Q,I IJ £ ... ~ Q,I \\ Q, x ...0 0.5 .J\ r~· Q, Q, .. \ ~ r----~l' i-+-.-+- 2 • 0.3 ·-·---· ..,. \D 00 N ..,. \D 0 N ~ 00 0 N ~ 00 0 N r- r- r- £ 00 00 00 ~ r- '°r- r- 00 00 00 00'° 00 ~ ~ ~ ~ ~ ~ ~ ~ °'~ °'~ ~ ~ ~ ~ ~ ~ ~ ·~ °'~ °'~

Annual Change in Arrivals

25 20 15 c 10 Q,I 5 ...IJ Q,I 0 Q, -5 -10 -15 -20 ..,. r- ~

FIGURE 2 Trend of arrivals from Japan and selected factors affecting them. Kawad and Prevedouros 23

The recent sharp recession in Japan tends to be confined to cer­ All the variables are statistically significant at the 95 percent level tain real estate and financial markets ("bubble, or overvalued econ­ (except for the intercept) and perform as expected. The GDPC has omy") and has not affected the majority of the public. To the con­ a positive impact on the arrivals whereas the CPI has a negative trary, various economic reports suggest that falling prices and effect. The dummy variable for strike has a negative impact, the per­ lowered inflation have made the recent recession a period of oppor­ formance of the stock market has a positive impact, and the effects tunity for most middle-class citizens of Japan. of Hurricane Iniki are negative, all as expected. The plot of the pre­ The variables used in the specifications of the Japan model are: dicted arrivals corresponds with the actual arrivals. The Durbin­ Watson statistic value of 1.9 indicates that autocorrelation is not a LLGNPC = Annual change of the GNP in percent; per­ threat to the validity of this model. The standard error of estimate is formed best in a logarithmic transformation, and 4.18, which is about half (51 percent) of the standard deviation of

lagged for 1 year (llGNPCN _ 1). the dependent variable. YEN = Dummy variable to account for large increases in thestrengthofthe¥: 1in1979, 1987, 1988, 1989, and 1992, 0 for all other years (taken from his­ The Japan Model torical futures charts of the Commodities Re­ search Bureau); performed best when lagged for The trend of arrivals shows that the arrivals from Japan increased one year (e.g., the large increase in 1978 affects moderately until the year 1986 (Figure 3, bottom). The years from the 1979 arrivals due to advance bookings). 1986 onward display a steep increase except for 1991, which shows WAR = Dummy variable to account for the Persian Gulf a decrease in the arrivals. The final model estimated with the war: 1 for 1991, 0 for all other years. Cochrane-Orcutt procedure is shown in Equation 2. RECESSION = Dummy variable to account for the recessions in 13 parameter t-statistic 1983 and 1991. %CHANGEJAPANN- 1 toN = - 4.28 [0.46] 175-79 = Dummy variable to account for the "cultural + 13.65 · LLGNPC [l.16] habit" of little vacation and foreign travel before +18.26·YEN [6.16] the 1980s. For this period, contrary to logical - 7.87 · WAR [1.28] expectations based on economic trends (see Fig­ -15.36 ·RECESSION [3.37] ure 2), foreign travel from Japan was low and - 5.65. fl5-79 [1.18] R 2 = 0.87, adjusted R 2 = 0.78, DW = 2.0 (2) roughly constant, although Japan's economy was growing vigorously. YEN and RECESSION are strongly significant, but the other vari­ ables demonstrate a significance at the 80 percent level only. All the Variables tested but excluded from the final specification included: variables perform as expected. It is evident that large fluctuations of the yen's strength affect the exchange rate accordingly, thus mak­ 1. The consumer price index at 1985 prices, and ing foreign travel more (or less) attractive to visitors from Japan. 2. The NIKKEI index (similar to the DJIA used for the USA The plot of the predicted arrivals corresponds with the actual market). arrivals. The Durbin-Watson statistic value of 2.03 indicates that autocorrelation is not a threat to the validity of this model. The stan­ The fit was not successful partly because NIKKEI 225 index sta­ dard error of estimate is 5.08, which is about half (54 percent) of the tistics before 1984 are not available. standard deviation of the dependent variable. Based on the models' ability to reflect the historical arrival pat­ terns, the following results summarize the estimation accuracy. The MODEL ESTIMATION worst 1- and 5-year underestimates for the USA model were -6.4 and -2 percent, respectively. The worst 1- and 5-year overesti­ The USA Model mates for the USA model were + 5. 7 and + 1.5 percent, respec­ tively. The error for the total period between 1975 and 1992 was The trend of arrivals shows that arrivals from the mainland United 0.046 percent. The worst 1- and 5-year underestimates for the Japan States increased for most years until 1991 (Figure 3, top). There was model were -7.5 and -1.2 percent, respectively. The worst 1- and a slight decrease from the years 1979 to 1981. An economic reces­ 5-year overestimates for the Japan model were+ 10.4 and +2.8 per­ sion in most parts of the country in late 1990 and early 1991 caused cent, respectively. The error for the total period between 1975 and a sharp decrease from the mainland United States in 1992. Addi­ 1992 was 0.046 percent. tional reasons for the decrease are the recession in California and the turbulence in the airline industry, which became a disincentive for travel to Hawaii because of the lowering of fares to most vaca­ Forecasts tion destinations, excluding Hawaii (14). The final model estimated with the Cochrane-Orcutt procedure is shown in Equation 1. Forecast models 1 and 2 can be made in the following two ways. 13 parameter t-statistic

%CHANGEUSAN- 1 toN = - 8.50 [1.23] 1. Short-term predictions of arrivals can be done using data pub­ + 30.82 · LLGDPC [2.86] lished by major organizations. Specifically, forecasts for the GDP - 2.17 ·CPI [4.61] and the CPI in the USA model can be input according to growth rates -16.09 ·STRIKE [3.54] reported by the OECD (15) and the EIU (16). The GNP growth rates -12.64 · NATURE [2.20] + 4.00 ·DJ/A [2.73] for the Japan model can be taken from the EIU. Population forecasts R2 =0.91,adjustedR2 =0.84,DW= 1.9 (1) for most countrie.s can be taken from the United Nations (17). 24 TRANSPORTATION RESEARCH RECORD 1506

4 en c: ~ 3.5 ~

3

2.5

2

1.5 1975 1980 1985 1990 1995 2000

I ARRIVALS FROM JAPAN

o+-~~~~...... --~~~~...... --~~~~-.--~---~~-.--~~~---; 1975 1980 1985 1990 1995 2000

FIGURE 3 Actual arrivals, forecasts, and confidence intervals for two major origins of travel to Hawaii, the mainland United States (top), and Japan (bottom).

2. For longer-term forecasts, the user must resort to a combina­ also depicted in Figure 3. These values were not included in the tion of trend extrapolation with ARIMA and educated estimates model estimation process and are shown here for comparison. based on contemporary macroeconomic literature basically The USA model could not account for the large decline in arrivals because, beyond two years, macroeconomic forecasts are erratically in 1993. A reversal of that drop in arrivals occurred last year and documented and often unavailable. The forecasts presented in Fig- · economic analysts expect travel to Hawaii to increase (J 8). Actual ure 3 are based entirely on ARIMA applications. After extensive arrivals are expected to return to the rates depicted by the forecast. trial-and-error specifications, the best-fit ARIMA models were The decline in arrivals, particularly in 1993, can be attributed USA-GDP ( 1, 1,0), USA-CPI ( 1,0,0), and Japan-GNP ( 1,0,0). Fore­ partly to a drop in airline seating capacity (lift) from the mainland cast values for the dummy variables were assigned so that the 1993- United States to Hawaii. Lift levels were approximately 7,600,000 to-2000 average matches the l 975-to-1992 average. in 1990, 7,200,000 in 1991, and 5,600,000 in 1993 (18). The corre­ sponding load factors were 63 percent in 1990, 63.5 percent in 1991, ARIMA modeling enables the estimation of confidence intervals. and 67.5 percent in 1993. Although supply clearly exceeds demand, The 90 percent-level confidence intervals are shown in Figure 3 for the increasing load factors and reduced number of scheduled flights the forecast portion of the models. The recent trend of arrivals from in peak seasons are likely to cause seat reservation difficulties for the mainland United States and Japan (years 1993 and 1994) are potential visitors, and ultimately the abandonment of trips to Hawaii Kawad and Prevedouros 25 for those who have little time flexibility. A reason for the reduction the state becomes more crowded, the hotels older, etc.). The model of lift levels is given in the Bank of Hawaii's Annual Economic also can be estimated using monthly arrivals to account for seasonal Report (18): "Much of the contraction of air service has arisen from effects. airline concern with 1.2 billion mi of unredeemed frequent-flier claims. Not surprisingly, frequent fliers are more apt to redeem their miles on travel to Hawaii than they are on flights to the destinations ACKNOWLEDGMENT on which the mileage was originally accumulated, so that Hawaii becomes an unprofitable destination for the carriers." The research for this paper was partly supported by the Airports A notable narrowing of the confidence intervals appears for 1994 Division, Hawaii DOT. and 1999 in the USA model. The reason is that the GDP and CPI have opposite signs for those years and, by coincidence, the errors largely canceled each other. REFERENCES The Japan model corresponds remarkably well with actual arrivals. However, the overall forecast may be somewhat optimistic. 1. Kawad, S. Aviation Forecasting at the Local Level: Theoretical Devel­ This is because the ARIMA forecast of the GNP assumes that the opment and Application at the Honolulu International Airport. M.S. trend of a rapidly growing Japanese economy will continue. Reports thesis. Department of Civil Engineering, University of Hawaii at Manoa, Honolulu, Hawaii, March 1994. in the financial press, however, state that Japan is less likely to sus­ 2. The Hawaii Econometric Model and its Applications. Department of tain as rapid a growth as in the past due to tightening worldwide Business, Economic Development and Tourism, Honolulu, Hawaii, competition, the easing of trade barriers, and a partial loss of the 1983. pricing advantage on several durable products. Thus, the near-term 3. Domestic Market Report (also Foreign Market Report). Hawaii Visitors Bureau, Honolulu, Hawaii, 1991. forecasts for Japan should be based on published GNP forecasts, 4. Revised Long-Range Economic and Population Projections to 2010: and the longer-term forecasts should be revised frequently. State of Hawaii (Series M-K). Department of Business and Economic Development, Preliminary Report, Honolulu, Hawaii, 1988. 5. Stuart, M. M. The Use of Gravity Model in Interpreting Air Passenger Traffic Between Hawaii and the Coterminous United States. Thesis. CONCLUSIONS Columbia University, New York, N.Y., 1968. 6. Crouch, G. I., L. Schultz, and P. Valeiro. Marketing International The goal of the research was the development of a short- to Tourism to Australia. Tourism Management, June 1992. medium-term econometric model system for forecasting air travel 7. Armstrong, C. W. G. International Tourism: Coming or Going, The Methodological Problems of Forecasting. Futures, June 1972. arrivals for the Hawaii DOT. The target use of the forecasts is land­ 8. Witt, S. F., and C. A. Martin. Econometric Models for Forecasting side planning applications. The model system is developed from Tourism Demand. Journal of Travel Research, Vol. 25, No. 3, 1987. readily available information, enabling the Airports Division to use 9. Arthur Young, Inc. A Report on Tourism Destinations Competing with it with no more effort than annual visits to the library to collect Hawaii. Department of Business and Economic Development, Hon­ olulu, Hawaii, 1989. updated data. Models for the United States and Japan have been 10. Liu, J.C., and J. Auyong. Tourist Attractiveness of Hawaii by County. developed using the Cochrane-Orcutt regression estimation proce­ Travel Industry Management, Tourism Research Publications, No. 10, dure. Similar models for Australia, Canada, Germany, Korea, New University of Hawaii, 1988. Zealand, and the United Kingdom were developed recently and are 11. Ritchie, J. R. B., and M. Zins. Culture as a Determinant of the Attrac­ presented elsewhere (19). tiveness of a Tourism Region. Annals of Tourism Research, April/June 1978. Various diagnostic tests were conducted to arrive at the final 12. Gee, C. Y., J. C. Makens, and D. J. L. Choy. The Travel Industry. Van models. All the explanatory variables perform as expected. The per­ Nostrand Reinhold, 1989. capita GDP (GNP for Japan) had a positive effect on arrivals, 13. Pacific Asia Travel Association (PATA). Annual Statistical Report, whereas the CPI had a negative effect. Large fluctuations of the I99I. 14. Bank of Hawaii. Annual Economic Report, Vol. 42, 1992. yen's strength affect arrivals from Japan. The effect of the airline 15. Organization for Economic Cooperation and Development. OECD Eco­ strike of 1985 was considerable. The effects of war (uncertainty, nomic Outlook, Paris, France, 1993. fear of terrorism, etc.) were shown to suppress foreign travel. The 16. Country Report. The Economist Intelligence Unit, London, 1994. coefficient estimates reflect that the Persian Gulf war had a nega­ 17. World Population Prospects, 1992 revision, United Nations, New York, tive impact on arrivals from Japan. The USA model shows that 1993. 18. Bank of Hawaii, Annual Economic Report, Vol. 44, Honolulu, Hawaii, ·arrivals also are sensitive to the trends in financial markets. 1994. The models correspond remarkably well to the actual number of 19. Prevedouros, P. D., R. Uwaine, and P. An. Origin-Specific Visitor arrivals, and all independent variables perform as expected. The Demand Forecasting at the Honolulu International Airport. Presented USA model has an overall error rate (for the 18 years of the esti­ at 7th World Conference on Transport Research, Sydney, Australia, July 1995. mation span) of less than 0.05 percent. Annual extreme values of error range from -6.4 to 5.7 percent. The Japan model also has an This paper reflects the views of the authors, who are responsible for the facts overall error rate of less than 0.05 percent. Annual extreme values and accuracy of the data presented herein. The paper does not necessarily of error range from -7 .5 to 10.4 percent. reflect the official views of the Airports Division, Hawaii DOT. This paper does not constitute a standard, specification, or regulation. Potential improvements to these models would entail the consid­ eration of other variables, such as the airline seat capacity to Hawaii Publication of this paper sponsored by Committee on Aviation Economics and the tourist attractiveness of Hawaii (which may be declining as and Forecasting. 26 TRANSPORTATION RESEARCH RECORD 1506

Flight Sequencing in Airport Hub Operations ·

CHING CHANG AND PAUL SCHONFELD

Airlines operating hub and spoke networks (HSNs) can reduce aircraft cost extreme solution as an initial solution and use a sequential pair­ costs and passenger transfer times at hubs through efficient sequenc­ wise exchange algorithm to improve that initial sequence until no ing of flights. Typically, batches of flights are processed during rela­ further improvement is possible. tively brief time "slots." When aircraft differ significantly in sizes or Our flight sequencing problem is to find a sequence that mini­ loads, there is a considerable potential for reducing the delay costs through efficient flight sequencing. Sequencing bigger aircraft last in mizes the total costs of passenger transfer delay, aircraft ground and first out (BLIFO) minimizes the costs of aircraft delays, gate usage, time, and gates. It is difficult to optimize exactly the sequence for a and passenger time. Sequencing smaller aircraft first in and first out batch of N arrivals and N departures at a hub because there are (N!)2 (SFIFO) maximizes the gate utilization and terminal capacity. There­ possible sequences. An efficient heuristic method to solve this flight fore, BLIFO is preferable when airports are not busy and gate utiliza­ sequencing problem is proposed in this report. tion is unimportant. SFIFO is preferable when airports are very busy. The literature on flight sequencing to minimize the .costs of the Some intermediate sequences might also minimize total cost, depend­ ing on the relative costs of aircraft delays, gates, and passenger time. passenger transfer delay, aircraft ground time, and gate use is BLIFO or SFIFO, whichever is lower, provides a very good initial scarce. Previous studies mostly focus on the Aircraft Sequencing solution in most cases. A sequential pairwise exchange algorithm can Problem (ASP). In each of these ASP models (1,2), a static prob­ then improve this initial sequence until no further improvement is lem is considered, in which N aircraft are already present on hold­ possible. ing stacks outside the terminal area. Each aircraft can land at any time, and the problem is to find the sequence that maximizes run­ Hub and spoke networks (HSNs) have been widely adopted by way capacity (or utilization) or, alternatively, minimizes delays. U.S. domestic airlines because they can greatly reduce the cost of Dear (1) examined the dynamic case of the ASP in which the com­ connecting a given number of cities and improve the service fre­ position of the set of aircraft varies over time. Psaraftis (3) devel­ quency. When compared with direct flights, the main disad­ oped a dynamic programming (DP) approach for sequencing N vantages of the HSN routing are additional transfer times and groups of aircraft landing at an airport to minimize total passenger costs at the hub. To minimize transfer times and costs, a batch of delays. Dear and Sherif_ (4) examined the constrained position aircraft has to arrive and depart within a short "time slot," and all shifting methodology, with simulation from the perspectives of aircraft should be on the ground simultaneously for at least a short both pilots and air traffic controllers, and later (5) developed a period so that transfers can be made. The common ground computer system to assist the sequencing and scheduling of termi­ time window (GTW) for passenger and baggage redistribution nal area operations. Bianco et al. (2) proposed a combinatorial opti­ is the time between arrival of the last flight and departure of the mization approach to the ASP, in which maximizing the runway first flight. The size of aircraft is related to passenger loads on dif­ capacity or utilization was modeled as an n job (landing or take­ ferent flights, especially for long-run scheduling purposes. Large off) and one-machine (runway) scheduling problem with non-zero aircraft imply expensive aircraft and large passenger loads. If ready times. Venkatakrishnan et al. (6) developed a statistical the sequencing allows larger aircraft to spend less time at the model for the landing time intervals between successive aircraft hub, the costs associated with the flight sequencing will be using data from Logan Airport in Boston. They found that reorder­ reduced. For instance, later arrivals and earlier departures for the ing the sequence of landing aircraft could substantially reduce the larger aircraft would reduce average passenger delay time and air­ landing time intervals and thereby increase runway capacity. Con­ craft ground time cost. Thus, total transfer passenger delay and air­ sidering stochastic aircraft arrivals, Hall and Chong (7) developed. craft cost would be reduced if larger aircraft were the last in and a model for scheduling flight arrivals and departures to minimize first out. delays for passengers connecting between aircraft at a hub Extreme sequences such as last in first out (LIFO), first in first out terminal. (FIFO), and their variants can be shown to minimize certain cost A review of the aforementioned studies indicates that determin­ factors. Under certain traffic conditions or when certain cost factors istic models for optimizing runway capacity or utilization have dominate sequencing decisions, it can be shown that particular received considerable attention. However, it is also important to extreme sequences are optimal. In more complex cases, where no consider the costs of passenger transfer delay, aircraft ground time, factor dominates and several factors must be traded off, sequencing and gate use at hub airports. In addition, departure sequences are solutions are also more complex. In such cases we will take the least interrelated with arrival sequences (for instance, due to minimum ground time constraints, and desirability of replacing departing air­ craft with similarly sized arriving aircraft to improve gate utiliza­ tion) and should be determined jointly. The flight sequencing prob­ C. Chang, Industrial Management Department, Chung-Hua Polytechnic Institute, Hsin Chu, Taiwan 30067, R.O.C. P. Schonfeld, Civil Engineering lem considered here is to find the arrival and departure sequences at Department, University of Maryland, College Park, Md. 20742. a hub that minimizes those three costs. Chang and Schonfeld 27

SYSTEM DEFINITIONS Cycle Time

A system of hub airports is defined as follows. The cycle time is the interval between the first arrival and the last departure in a batch of flights, along with the buffer separation time (Figure 3). The cycle time has four components: Route Network 1. The arrival period is the sum of all interarrival times, which is An HSN that has one hub airport and N spoke city airports (Figure l 7=-,' A;, where N is the total number of aircraft, and A; is the inter­ 1) is considered here. All spoke routes connect at the hub. To travel arrival time between the ith and i + 1st arrivals. from one spoke city to another, passengers must transfer at the hub. 2. The GTW is the common time when all aircraft are simulta­ Nonstop travel is possible only if the origin or destination is at the neously at the terminal, for transfer purposes. (However, transfer hub. A more general system would have multiple hubs. activities can start before the GTW and continue after its end.) 3. The departure period is the sum of all interdeparture times, which isl 7~ 1 D;, where D; is the interdeparture time between the ith Batch Arrivals and Batch Departures 1 and i + 1st departures. 4. Buffer separation time, q, is the minimum separation time A group of flights from various spoke cities arrive at the hub airport between successive slots, which is constrained by reliability con­ within a short time period, and then unload and load passengers and siderations. baggage during a common GTW; then, the same aircraft leave within a short departure period. If there are N arriving aircraft, there The first three time components are available for passengers, bag­ are also N departing aircraft. gage, and cargo transfer activities. Aircraft are ready for departure after loading and servicing processes are completed. Sequence

The sequence is the order of aircraft arrivals and departures. Two Slot Sequences extreme sequences, FIFO and LIFO, are of special interest if aircraft are ranked by size or load. FIFO is the sequence in which aircraft Here, the time between the first aircraft arrivals of two consecutive depart in their order of arrival [Figure 2(a)]. LIFO is the sequence batches is called a time slot. Figure 3(a) shows that if cycles do not in which aircraft depart in their reverse arrival order [Figure 2(b)]. overlap, the slot duration equals the cycle time. However, if cycles LIFO is interesting because it may allow larger aircraft and their overlap [Figure 3(b)], the slot duration is smaller than the cycle passengers to arrive later and leave earlier, with considerable sav­ time. ings. FIFO is interesting because it can reduce slot durations. At Figure 4 shows two types of slot sequences. busy airports in which gate utilization and terminal capacity are crit­ ical, a FIFO sequence can provide shorter intervals among succes­ 1. Overlapping cycles are possible if the departure sequence in sive batches of flights, as discussed later in this paper. Two extreme the leading slot is similar to the arrival sequence in the trailing slot, FIFO sequences are considered in which the aircraft order is by size. as when: These are SFIFO, in which smaller aircraft are first, and BFIFO, in a. All slots are SFIFO or BFIFO [Figure 4(a)], which bigger aircraft are first. Likewise, the LIFO options include b. Any pair of successive slots includes one SLIFO and SLIFO, in which smaller aircraft arrive last and depart first, and BLIFO [Figure 4(b)]. BLIFO, in which bigger aircraft arrive last and depart first. 2. Other nonoverlapping cycles can have a. Random sequences, b. Alternating SFIFO and BFIFO slots [Figure 4(a)], c. Similar LIFO sequences; that is, all SLIFO or all BLIFO ··ay slots [Figure 4(a)].

COST FUNCTIONS !/~, ;~\ \ Three cost components reflect the effects of different batch sequences. These are the total passenger transfer delay cost ( Cp), the

), ). total aircraft ground time cost ( C0 and the total gate cost ( C8 Local passengers (originating or terminating at the hub) can be \~®~CD... excluded in these total cost functions because there is no difference \®~ / between HSN routing and direct flights. All these costs are com­ puted in the same units ($/slot). The total relevant cost function is

(1) ········...... ········· ·· ... ·· .... at ...... · where C = total cost per slot for the hub operation ($/slot). FIGURE 1 Hub and spoke network. We define these three component costs as follows. 28 TRANSPORTATION RESEARCH RECORD 1506

# of passengers

GTW

Time

(a) FIFO Sequence

# of passengers

GTW

Time (b) LIFO Sequence

FIGURE 2 FIFO and LIFO sequences.

a. The passenger transfer delay is incurred by redistributing period so that transfers can be made. An aircraft's ground time is transfer passengers and their baggage from their original aircraft to determined by its arriving and departing times. The total aircraft their destination aircraft at the hub. The delay for each passenger is ground time cost is the sum of ground time cost for all aircraft the difference between the passenger's departure time from the hub and the passenger's arrival time at the hub. Therefore, the total pas­ N Ca= G;Va; (3) senger transfer delay cost is the sum of delay costs for all transfer I i=I passengers in a batch of arrival and departure flights where

(2) Ca = total ground time cost for all aircraft in a slot ($/slot), a; = the time aircraft i is on the ground (hours/slot), and

v ; = ground time value of aircraft i ($/hour). where 0 c. The total gate cost includes the hourly gate fixed cost and cp = total passenger transfer delay cost per slot ($7slot), hourly gate usage cost. The fixed cost accounts for the gate con­ v" = time value of transfer passengers ($/passenger hour), struction and equipment installation. The usage cost is incurred Pu = number of transfer passengers from arrival aircraft i to when an aircraft parks and uses a gate. Because gates can have dif­ departure aircraftj (passengers), and ferent characteristics in the same terminal, the gate fixed cost tu = the transfer delay time from arrival aircraft i to departure depends on the gate size and slot duration, and the gate usage cost aircraft j (hours/slot). depends on gate size and gate occupancy time b. The aircraft ground time is the time that an aircraft dwells at N the hub. For HSNs, a batch of aircraft arrive and depart within a slot C8 =I (Ve; f8; + V 0 ; f 0 ;) (4) and all aircraft are on the ground simultaneously for at least a short i=l Chang and Schonfeld 29

GTW GTW q q /1 ~ ! ) A D /1A L\D Time Slot 1 Time Slot 2

Cycle Length 1 Cycle Length 2

(a) Non-Overlapping Cycles

Cycle Length 1

Time Slot 1 Time Slot 2

Cycle Length 2

(b) Overlapping Cycles

FIGURE 3 Overlapping and nonoverlapping cycles.

where CONSTANT SLOT DURATION

Cg = total gate cost per slot ($/slot), A simplified case with constant slot duration is considered first, on vc; = fixed cost of gate i ($/hour), the basis of these assumptions:

v0 ; = usage cost of gate i ($/hour),

t0 ; = gate i's occupancy time (hours/slot), and 1. All transfer passengers considered arrive and depart within the tg; = slot duration (hours/slot). same slot, and all transfer activities are completed within that slot.

SFIFO SFIFO BFIFO BFIFO

(a) Cycle Overlap (b) No Overlap (c) Cycle Overlap

BLIFO BLIFO ~s//__._1~_GTW__.__B~-"t---¥'-0---'-B~G_TW__,_BR~~I GTW 1)1• \:Eit q SSS S S S ~ ~ SLIFO SLIFO (a) No Overlap (b) Cycle Overlap (c) No Overlap

FIGURE 4 (a) Possible FIFO sequences; (b) LIFO sequences. 30 TRANSPORTATION RESEARCH RECORD 1506

2. The number of gates, G, is greater than or equal to the num­ ger transfer delay cost. To minimize the total passenger transfer ber of aircraft, N. delay, aircraft should arrive in ascending order of their passenger 3. The ground time window (T) is a constant. arrival rates (the number of arriving passengers/runway time unit) 4. Aircraft arrive punctually. and depart in descending order of their passenger departure rates. Because larger aircraft may need more ground time than smaller No other scheduling constraints limiting flight arrival and depar­ ones, the BLIFO departure sequence must be modified to consider ture times should be considered in this problem. The ground activi­ which aircraft are actually ready to leave. Smaller aircraft that are ties of an aircraft include unloading passengers, baggage, and cargo; ready early need not wait for the unready larger aircraft. Accord­ and cleaning, refueling, and loading passengers, baggage, and cargo. ingly, the aircraft departure sequence can be modified as follows. A, D, and Tare fixed quantities and are assumed in this case to be Step 1. Check all aircraft to find which ones are ready to leave. independent of the sizes of aircraft. The buffer separation time Step 2. Sort all ready aircraft in descending order of their pas­ between two time slots is q. Therefore, the slot duratio~ is constant. senger departure rates, and let the aircraft with the largest passen­ ger departure rate leave. Analysis of Sequences Step 3. Check the unready aircraft. If new aircraft become ready to leave, let them join the list of ready aircraft. Go to Step 2. We first explore the extreme sequences LIFO and FIFO to deter­ The way in which BLIFO minimizes total aircraft ground time mine in what situations they actually yield optimal solutions and cost can also be explained graphically. An aircraft's dwell time is then consider how more complex cases can be solved. the interval between its arrival time and departure time. Figure 5 shows that with BLIFO, larger aircraft have smaller dwell times.

Since for BLIFO a 1 :5 a 2 :5 ... :5 aN and Vai 2::: Va2 2::: ••• 2::: VaN' LIFO Sequence where a; is the ith largest aircraft dwell time and Va; is the ith largest aircraft's time value, BLIFO minimizes the total aircraft ground In the LIFO sequence aircraft depart in their reverse order of arrival. time cost (Equation 3). Thus, a LIFO sequence benefits later arrivals and makes earlier A similar argument can be used to show that BLIFO minimizes arrivals a disadvantage. If larger aircraft (with higher costs per air­ total gate usage cost since in Equations 3 and 4 the total gate usage craft hour and passenger loads) are required to arrive later than cost function has the same structure as the total aircraft ground time smaller aircraft, such a BLIFO sequence minimizes the total pas­ cost function. Consequently, BLIFO minimizes total passenger senger transfer delay cost, aircraft ground time cost, and gate usage transfer delay cost, total aircraft ground cost, and total gate usage cost. Figure 5 shows how BLIFO minimizes the total passenger cost, but not the total fixed cost of gates. transfer delay cost; the abscissa is time, and the ordinate is the cumulative number of passengers. The areas covered by the arrival curve, the (GTW), and departure curve in Figure 5 represent the FIFO Sequence total passenger transfer delay. Because the GTW has a fixed value, we only need to consider the areas covered by the arrival and depar­ When gates differ in size and cannot all accommodate the largest ture curves. The slopes of arrival .or departure curves represent the aircraft, the sequence of flights depends on the order in which gates number of arriving or departing passengers per time unit; that is, the of different sizes become available after handling the previous batch passenger departure rate. It can be observed in Figure 5 that the area of aircraft. With the gate-aircraft size compatibility restriction, a under the BLIFO arrival and departure curves is smaller than areas BLIFO slot cannot closely follow a preceding BLIFO slot. This for any other sequences. Thus, BLIFO minimizes the total passen- reduces the gate utilization and terminal capacity, which are very

#passengers

Arrival Departure

!$0,_1 ______?------~~ ~~% ______,_,.;._ __8:.__ .....------" ~

....1/ ____ !v,_;L , --- I ,

.,_...... , ______...... A T D Time

FIGURE 5 BLIFO and SLIFO sequences. Chang and Schonfeld 31

# of passengers • Arrival GTW Departure Arrival GTW Departure

.._Q_ ... -1!-- --=::JfQ! Q! r-~- 1- ,, --7 '7 Q:II I LJ1 L;2 ~_J Q! T T I 1: E2 Q! I

Q: ~I' Q:

.. ,. ti -A q A Time FIGURE 6 Overlapping slot sequences.

important at busy airports. To maximize the gate utilization and ter­ and I:n = Q:,,, for all m. However, if 1:n + Q:,,, for all m, neither minal capacity, the slots must overlap tightly. Two succeeding slots SFIFO nor BFIFO guarantees the minimum total passenger transfer can overlap tightly if the departure sequence in the leading slot is delay. similar to the arrival sequence in the trailing slot (Figure 6). FIFO Similarly, it is easy to find a sequence that minimizes total air­ yields tightly overlapping sequences for successive slots· when the craft ground cost and gate usage cost since both costs are related to departure sequence in the earlier slot is the same as the arrival aircraft sizes and their dwell times. Figure 6 shows that the follow­ sequence in the later slot. Thus, FIFO can increase gate utilization ing properties are true when slots must overlap tightly. (It should be and terminal capacity. Two extreme cases of FIFO, namely SFIFO noted that here I:n need not be equal to Q:,,, for all m, since both costs and. BFIFO, significantly affect the total passenger transfer delay are not related to the passenger loads.): when slots must overlap tightly. To minimize the total passenger d. If A> D, SFIFO minimizes total aircraft ground time cost and transfer delay, the areas of Z, and £ 1 in Figure 6, where t is the slot number, should be minimized. When interarrival times (A) and total gate usage cost. interdeparture times (D) have fixed values, the least transfer delay e. If A< D, BFIFO minimizes total aircraft ground time cost and total gate usage cost. sequence minimizes areas (£1 + Z1) in Figure 6, where t = 1. For instance, assume that there are five aircraft in each slot in Figure 6. f If A = D, all FIFO sequences have the same total aircraft Equation 7 represents the total passenger transfer delay. In mini­ ground time costs and total gate usage costs. mizing total delay, subject to the overlapping slot constraint, the fol­ For this simple case when slots must overlap tightly, either SFIFO lowing results are obtained: or BFIFO minimizes total aircraft ground time cost and gate usage cost. Area E, = 4AI[ + 3Ag + 2AI~ + Ati (5) When slots overlap tightly and I~, = Q:n, for all m, the least total cost sequence is: Area Z, = Dm + 2Dm + 3DQ~ + 4DQ~ (6) g. SFIFO, if A > D. h. BFIFO, if A < D. Min Area (E, + Z,) = 4A/[ + (3AI~ + DQ2) + 2AI~ + 2DQD i. All FIFO sequences have the same total costs, if A = D. +AI~+ 3DQD + 4Drn (7) If l!n + Q:,,, for all m, the above results may not be true. However, where (d), (e), and if) are still true when slots must overlap tightly. If nei­ ther SFIFO nor BFIFO minimizes total passenger transfer delay, the /,~ = the total number of transfer passengers on the mth arrival sequence which minimizes total passenger transfer delay has higher aircraft in slot t, and total costs of aircraft ground time and gate usage. Therefore, the Q,;, = the total number of transfer passengers on the mth depar­ total cost of SFIFO, if A> D, or BFIFO if A< D, is very close to ture aircraft in slot t. the minimum total cost when slots must overlap tightly (8). If/,;, = Q,;, (the number of the transfers on mth arrival aircraft is similar to the number of the transfers on the mth departure aircraft), 3.2. Preferable Sequence for all m, the following is true: When an airport is not busy and the gate utilization (gate fixed cost) a. If A > D => {fl < l2 {/[>/~.>I~> I~>/~} and {Ql >Qi> Qj > If aircraft are not ready to leave as soon as BLIFO sequence m> QD minimizes areas of (E, + Z,). This sequence is BFIFO .. requires, the BLIFO departure sequence should be modified as in c. If A = D =>all FIFO sequences have the same transfer delay. the LIFO sequence already described. When an airport is very busy and slots must overlap tightly, a Accordingly, either the SFIFO or BFIFO flight sequence mini­ FIFO sequence (specifically SFIFO if A> D and BFIFO otherwise) mizes total passenger transfer delay when slots must overlap tightly maximizes gate utilization as well as terminal capacity and is the 32 TRANSPORTATION RESEARCH RECORD 1506 least total cost sequence if 1:,, = Q,~,, for all m. When I:,, + Q:m for a. AiJ ~ Aji, if i ~ j, all m, neither SFIFO nor BFIFO may be the least total cost over­ b. DiJ ~ Dj;. if i ~j, lapping sequence. However, either SFIFO or BFIFO is still prefer­ c. AiJ ~ DiJ, for all ij pairs. able because the total cost of SFIFO or BFIFO is very close to the minimum total cost. When an airport's condition is moderately busy, trade-offs Overlapping Sequence among passenger time, aircraft costs, and gate cost may lead to a least total cost sequence in between extreme sequences such as In order to maximize the gate utilization and terminal capacity, slots BLIFO, BFIFO, or SFIFO. Moreover, the time values of pas­ should overlap tightly. When interarrival times and interdeparture sengers, aircraft, and gates vary in different times and places. In times are dependent on the relative weight classes of two successive such cases, the least total cost sequence may be found by starting landing and takeoff aircraft, FIFO can still increase the gate utiliza­ from some initial solution and using the sequential pairwise tion. Assume that departure processes are fixed. Based on these exchange algorithm to try swapping aircraft positions in the properties, if aircraft arrive in the order of {Small, Large, and sequence until no further improvement is possible. We can choose Heavy}, the minimum total arrival time is obtained. Due to prop­ the best extreme solution (i.e., BLIFO or SFIFO if A > D and erty (a), Au would be smaller than Aji if i ~ j. In order to minimize BFIFO otherwise) as our initial solution and then improve it with total arrival time, small aircraft should land before large aircraft. a systematic exchange algorithm. The total number of exchanges Similarly, for a. fixed arrival process, in order to minimize total is N(N - 1)/2, where N is the total number of aircraft, for example, departure time, smaller aircraft should take off before larger air­ [1,2], [1,3], ... , [1,N], [2,3], [2,4], ... , [N - 2,N - 1], [N - craft. Accordingly, SFIFO minimizes cycle length and slot duration 2,N], [N - l,N]. For instance, assume that A > D. Our sequential since SFIFO has the smallest total arrival and departure times. pairwise exchange algorithm to improve the flight sequencing is as Therefore, when slots must overlap tightly, SFIFO is the overlap­ follows. ping slot sequence that maximizes the gate utilization and terminal Step 1. Compute the total costs of SFIFO and BLIFO. The one capacity. with the lower total cost is the initial solution. Store its total cost. Because the interdeparture time is slightly shorter than the inter­ Step 2. Sequentially choose a pair of aircraft and exchange their arri val time, SFIFO benefits larger aircraft. This implies that SFIFO arrival orders. Compute the new total cost. is the overlapping sequence that minimizes the total cost of aircraft Step 3. If the new total cost is below the previous one, substitute ground time. it and store the new arrival sequence. Otherwise, keep the previous SFIFO has the smallest gate time and can minimize total gate sequence. Go to Step 2. fixed cost because the shortest slot duration sequence yields the This algorithm was used by Chang (8) and had a reasonable com­ highest gate utilization. In addition, SFIFO minimizes the total air­ putation time. craft ground time. Therefore, SFIFO also minimizes total gate usage cost. Consequently, SFIFO is the overlapping sequence with the least total gate cost. On the basis of the results of the constant slot duration case, if VARIABLE SLOT DURATION interarrival time, AiJ, is greater than interdeparture time, Du, for all i and}, and 1:,, = Q:,, (the number of the transfers on mth arrival air­ When the interarrival and interdeparture times are variable and the craft is similar to the number of the transfers on the mth departure GTW is constant, slot duration differs for various flight sequences. aircraft), for all m, SFIFO is the overlapping sequence with the least If.an airline accounts for a significant fraction of the flights at an air­ total passenger transfer delay. port, the runway capacity directly affects the interarrival and inter­ With SFIFO, similarly sized aircraft arrive or depart together, departure times of an airline's batch of connecting flights. One key consistent with the principle of grouping takeoffs and landings of factor that can affect the interarrival and interdeparture times is the similarly sized aircraft (6). Similarly sized aircraft land or take off minimum separation required by FAA to guard against wake-vortex together and average interarrival time and interdeparture time are turbulence (9). The wake-vortex separation depends on weights of minimized. Therefore, SFIFO maximizes runway capacity in such the leading and following aircraft. Three weight classes of aircraft hub operations. (heavy, large, and small) must be considered. Thus, SFIFO is the least total cost overlapping sequence if /,~, = Q/,,, for all m. Otherwise, SFIFO may not minimize total pas­ senger transfer delay. The SFIFO Sequence section of this paper Minimum Separation Requirement also indicates that the total passenger transfer delay of SFIFO is close' to the optimal value. Moreover, SFIFO still minimizes total Let Au be the interarrival time between two successive landing air­ aircraft ground time cost and total gate cost (including gate usage craft i and j, and Du be the interdeparture time between two suc­ and gate fixed costs) when slots must overlap tightly. Thus, SFIFO cessive take-off aircraft i and}, where both i and} are aircraft size is a near-optimal overlapping sequence because its total cost is close indices. Aircraft are ordered and labeled according to decreasing to the minimum total cost when slots must overlap tightly. size; for example, { 1, 2, 3, 4} are heavy aircraft, {5, 6, 7, ... , 10} are large aircraft, and { 11, 12, ... , N} are small aircraft. Let the time period between the first arrival and the last arrival be called 4.3. Nonoverlapping Sequence total arrival time, and the time period between the first departure and last departure be called total departure- time. Based on the When an airport is not busy and gate utilization is unimportant, FAA's minimum separation regulation, the following properties BLIFO is still preferable. BLIFO is the sequence in which aircraft exist: arrive in ascending order of their passenger arrival rates (the num- Chang and Schonfeld 33 ber of arriving passengers per runway time unit) and depart in SFIFO does not mm1m1ze total passenger transfer delay, the descending order of their passenger departure rates. This always sequence that minimizes total passenger transfer delay has higher benefits large aircraft and reduces total cost significantly. However, total costs of aircraft ground time and gate usage because SFIFO BLIFO may have a longer slot duration than SFIFO. The total inter­ minimizes these two costs. Besides, total passenger transfer delay arrival time for BLIFO is the same as that for SFIFO, but the total of SFIFO is close to the optimal value. Without 1,~ = Q{n, for all m, interdeparture time for BLIFO is greater than that for SFIFO. For SFIFO may not be the optimal overlapping sequence but is still instance, assume { 1, 2, 3,4} are heavy aircraft, { 5, 6, 7, 8} are large near-optimal. aircraft, and {9, iO} are smaii aircraft. Let the departure sequence When an airport is moderately busy, neither BLIFO nor SFIFO of SFIFO be (10, 9, 8, 7, 6, 5, 4, 3, 2, 1). The departure sequence of may be the optimal sequence. In addition, the time values of pas­ BLIFO is (l, 2, 3, 4, 5, 6, 7, 8, 9, 10). Therefore, the difference of sengers, aircraft, and gates vary in different times and places. As the the slot durations between BLIFO and SFIFO is the difference of time value of the gates increases relative to other costs, SFIFO is the separations, that is, ur =I (D;,; +I - D;+u) = D45 - Ds4 + Ds9 - increasingly preferable to BLIFO. To find the optimal sequence,

D 98 , divided by the take-off speed. Other interdeparture times are BLIFO or SFIFO, whichever has the lower total cost, is used to be the same (e.g., D, 2 = D21 ) since interdeparture times for SFIFO and the initial solution and improved by the sequential pairwise BLIFO are equal if two successive takeoff aircraft are in the same exchange algorithm until no further improvement is possible. How­ weight class. For instance, if all aircraft are in the same weight class, ever, this improved sequence may not be the exact optimal sequence BLIFO and SFIFO have the same slot duration. Since FAA defines since the flight sequencing problem is an NP-hard problem (2) when only three weight classes and BLIFO is also consistent with the an airport is moderately busy. principle ot grouping landings or takeoffs of similarly sized aircraft, In this report, the GTW is assumed to be independent of flight the difference in total departure times between SFIFO and BLIFO sequence, even though the minimum ground times of smaller and is small. This difference can be ignored if the number of aircraft in larger aircraft are considered. Improved models should explicitly con­ a slot is large. BLIFO has a very small cycle time. The arguments sider how the GTW is affected by a flight sequencing. In addition, the used in the section LIFO Sequence can also be used to show that flight sequencing and gate assignment are interdependent. In a previ­ BLIFO with variable interarriv.al and interdeparture times mini­ ous report Chang (8) has analyzed a more realistic flight sequencing mizes the costs of total passenger transfer delay, total aircraft problem with a variable ground time window and combined gate ground time, and total gate usage. The BLIFO departure sequence assignment, and also has provided extensive numerical results. can again be modified to deal with unready flights with the proce­ dures described in the LIFO Sequence section. REFERENCES CONCLUSIONS l. Dear, R.G. The Dynamic Scheduling of Aircraft in the Near Terminal The flight-sequencing problem considered here is to seek an effi­ Area. MIT Flight Transportation Laboratory Report R-76-9, MIT, Cam­ cient flight sequence that minimizes the total costs of passenger bridge, Mass., 1976. 2. Bianco, L., G. Rinaldi, and A. Sassano. A Combinatorial Optimization transfer delay, total aircraft ground time, and gates. When aircraft Approach to Aircraft Sequencing Problem. Proc., NATO Advanced differ significantly in size or load, there is considerable potential Research Workshop on Flow Control of Congested Networks, Capri, for reducing the costs through efficient flight sequencing. In addi­ Italy, October 12-18, 1986, pp. 323-339. tion, aircraft landings and takeoffs must satisfy the minimum sepa­ 3. Psaraftis, H.N. A Dynamic Programming Approach for Sequencing ration requirement. The interarrival times and interdeparture times Groups of Identical Jobs. Operations Research, Vol. 28, No. 6, 1980, pp. 1247-1359. depend on the weight classes of two successive aircraft landings 4. Dear, R.G., and Y.S. Sherif. The Dynamic Scheduling of Aircraft in High or takeoffs. The flight-sequencing disciplines that favor large air­ Density Terminal Area. Microelectronics and Reliability, Vol. 29, No. 5, craft such as SFIFO and BLIFO may minimize the considered total 1989, pp. 743-749. cost under some circumstances. Even if SFIFO or BLIFO does 5. Dear, R.G., and Y.S. Sherif. An Algorithm for Computer Assisted Sequencing and Scheduling of Terminal Area Operations. Transporta­ not minimize the total cost, one of them (the one with the lower tion Research, Part A, Vol. 25, Nos. 2/3, 1991, pp. 129-139. total cost) will be a good initial solution for the flight sequence, 6. Venkatakrishnan, C. S., A. Barnett, and A.R. Odoni. Landings at Logan which can then be improved with the sequential pairwise exchange Airport: Describing and Increasing Airport Capacity. Transportation Sci­ algorithm. ence, Vol. 27, No. 3, 1993, pp. 211-227. When an airport is not busy, the gate utilization is less important 7. Hall, R.W., and C. Chong. Scheduling Timed Transfers at Hub Termi­ nals. Transportation and Traffic Theory (C.F. Daganzo, ed.), 1993, pp. and gate-fixed cost can be neglected. BLIFO is then preferable since 217-236. it minimizes the costs of total passenger transfer delay, total aircraft 8. Chang, C. Flight Sequencing and Gate Assignment at Airport Hubs. ground time, and total gate usage. When an airport becomes busy, Ph.D. Dissertation, University of Maryland at College Park, 1994. the gate utilization and terminal capacity become more critical and 9. 1991-92 Aviation System Capacity Plan, Federal Aviation Administra­ tion, Washington, D.C. slots should overlap tightly. SFIFO is the least total cost overlap­ ping sequence if 1,;, = Q,;,, for all m. However, if 1:,, + Q,;,, for all m, Publication of this report sponsored by Committee on Aviation Economics then SFIFO may not minimize total passenger transfer delay. When and Forecasting. 34 TRANSPORTATION RESEARCH RECORD 1506

An Optimum Resource Utilization Plan for Airport Passenger Terminal Buildings

MAHMOUD S. PARIZI AND JOHN P. BRAAKSMA

An ideal and practical procedure was developed to optimize resource primarily caused by airlines concentrating flights at certain times of utilization of airport Passenger Terminal Buildings (PTBs). Each pro­ the day, days of the week, or seasons of the year. The airline sched­ cedure consists of three parts, that is, the PTB operation, an optimiza­ ule is established based on several criteria, such as the public tion model, and a flow management and control model. In the ideal pro­ demand to travel within social hours, the need for arranging con­ cedure, it is assumed that a real-time flow management and control technique can be applied on an actual terminal building to dynamically necting services, the utilization pattern of aircraft fleet, and other allocate the optimum required resources, obtained from the optimiza­ constraints, for example, night curfews, numerical limitation to tion model, to a highly variable demand. The output of this procedure night movements, or noise limitations. would be a variable time-resource plan and theoretical optimum opera­ The second factor is the current design procedure for PTBs, tions cost. In the practical procedure, the PTB is simulated using a new which determines space requirements according to a broad criterion object-oriented graphical modelling technique, the SES/workbench. of average space per person (2). For a given traffic level, there is a The object and submodel support of this tool allowed rapid develop­ ment of the simulation model capable of representing a wide variety of corresponding space requirement. The traffic level is the forecast PTBs. Statistics from the simulation model are used to develop an opti­ demand for a typical hour, that is, typical peak hour passenger mization model, based on the resource allocation theory, to yield the (TPHP). Derivation of the TPHP from the existing or forecast data optimum required resources for each segment of the building at each varies among countries, and there is no universally acceptable def­ instant of time. It is also proposed that the results of the optimization inition for the TPHP. Even with the standard TPHP, the problem model, a variable time-resource plan, can be implemented by applica­ still remains. In theory this is because, assuming the average peak tion of some site-specific flow management and control strategies. As a result, the procedure will provide a practical variable operational and as the 30th busiest hour, the PTB is fully utilized .for 1 hour in the maintenance cost. How close one can bring the practical cost to the the­ year, overutilized for 29 hours in the year, and underutilized for the oretical one depends on the flexibility of the PTB layout and the capa­ reminder of the time (8730 hours; there are 8760 hours in a 365-day bility of flow management and control technique. year) (3). In other words, the PTB is overutilized to fully utilized only 0.34 percent of the time each year. The third factor is that the total system cost is often not visible, Capital and operational costs of airport passenger terminal build­ particularly those costs associated with operation and support. The ings (PTBs) are very extensive. Taking into account the fact that air­ visibility problem can be related to the iceberg effect, in which the port PTBs are some of the most expensive public transportation visible parts are design and construction costs and the remainder facilities, the goal of public policy should be to ensure that these resources are employed as effectively as possible. A quick review (under the water) are operating and maintenance costs. Therefore, of the existing literature gave the impression that for most of the the operating cost of PTBs has been almost always ignored in the planning horizon, airport terminals have an oversupply of facilities, planning and design process. for example, space. However, various interest groups are concerned The fourth factor is that, in spite of a sensitive relationship about the negative impacts of insufficient supply during certain time between physical planning and operational planning, the analysis of periods, for example, congestion and delays (1). Considering that these two is not done early enough. To increase the efficiency of the there are undesirable consequences of oversupply as well as under­ PTB, two main concepts should be considered in concert: design supply, it is very important to size and operate airport facilities as and operation ( 4). The result of ignoring one of these two concepts realistically as possible. Although it may be difficult to design and would be an inefficient terminal. The combination of a poorly construct PTBs according to the variable nature of their demand, it designed PTB with an excellent operational plan may operate well, is possible to operate them more effectively. This is doubly impor­ whereas an excellently designed PTB with a poor operating plan tant with respect to the growth of operating and maintenance costs may result in a poorly operating terminal. as the PTB grows in scale and complexity. Factors that contribute The fifth factor associated with the airport PTB desigri and oper­ to the problem of oversupply and undersupply are summarized. ation is the month-to-month uncertainty of the airline industry. This creates situations in which terminals designed for one type of oper­ ation are forced to operate under a completely different situation FACTORS CAUSING THE PROBLEM because of the bankruptcy of a major airline, and situations in which one carrier is replaced by another carrier with very different sched­ The most obvious factor causing the ineffective utilization of air­ uling or types of passengers, that is, international versus domestic port terminals is traffic peaking. Peaking may be by hour of the day, or hub versus origin and destination operations (5). This uncertainty day of the week, and season of the year. Airport peaking is of course can sometimes render even well-conceived designs inefficient or inappropriate. Department of Civil & Environmental Engineering, Carleton University, However, with the costs and difficulties in modifying the infra­ Ottawa, Canada, KlS 5B6. structure to keep pace with the air traffic changes, it would be nee- Parizi and Braaksma 35 essary for airport planners to (a) favor a flexible design, (b) prepare Thus, based on the concepts of resource allocation, flow manage­ an operational plan for different scenarios, and (c) allocate the ment and control, and the dynamic function of the PTB, an ideal resources based on the variable nature of demand. One approach to optimization procedure can be developed. The ideal methodology this problem of misallocation of resources is explored in this paper of the proposed optimization procedure is indicated in Figure 1. The by investigating the application of resource allocation and flow process consists of three basic parts, that is, the operation of PTB, management and control theories to the operation of the PTBs. The an optimization model, and a flow management and control model. hypothesis holds that by applying a flow management and control Within the operational process, passengers will pass through the tool, the resources can be allocated to the variable demand in such PTB according to their prespecified schedules and some operational a way as to minimize the operational and maintenance cost of the guidelines. The optimum value of required resources would be PTB from an airport authority's point of view. found based on the demand placed on the system and performance measures by using an optimization model. These values would be allocated to different segments of the PTB that perform different IDEAL OPTIMIZATION PROCEDURE activities based on some flow management and control techniques. If the resources were allocated as they were found from the opti­ The PTB is a collection of components to facilitate the transfer of mization model, then the output would be minimum operational and passengers and their baggage from groundside to the airside, or vice maintenance costs for the PTB at different levels of performance. In versa, and sometimes between airsides. These components are con­ the following sections, the theoretical optimization model is for­ sidered as existing resources that are to be utilized based on the traf­ mulated, a more practical optimization procedure is develope.d, and fic demand. Since the demand on the terminal system is stochastic a new object-oriented simulation model is discussed to take the and variable, the utilization of resources should also be variable. place of a real terminal building.

Daily Demand

Pax. Terminal Operational Buildin Guidelines

Measures of Performance

Optimization Model

Optimum Required Resources

l+------1 Flow Management and Control Techniques

Allocation of Scarce Resources

PTB With Optimum Operational and Maintenance Cost For the Desired Level of Performance

FIGURE 1 Ideal procedure of optimum PTB resource utilization. 36 TRANSPORTATION RESEARCH RECORD 1506

Theoretical Analysis process and assuming ~j as the unit cost of undersupply, the expected undersupply cost would be: As mentioned earlier, the PTB is a dynamic system mainly consist­ ing of two types of entities, that is, resources and objects (passen­ (4) gers). To optimize the utility of the system, one may deal with either resources or objects or both. In this research, resource allocation where Y =maximum expected demand for segment}. theory is used as the basis of the optimization model. In resource Therefore, the total over and undersupply cost associated with the allocation theory a fixed amount of resources are allocated to a allocation of x resources to segment j is the sum of two preceding series of activities with variable demand in such a way that the 1 cost elements as follows: objective function under consideration is optimized. The resource allocation problem is generally formulated as follows: (5)

Minimize J (xi. X2, ... Xn) where Subject to: Li=i x1 = N, x1 > O,j = 1, 2, 3 ... , n (1) c/x1) = expected cost at segment}, That is, given one type of resource, for example, space whose total x1 = resource for allocation, amount is equal to N, we want to allocate it ton service locations p/y) = probability mass function of demand, (segments of PTB) which serve an uncertain number of customers 81 = resource allocated to each demand unit, so that the objective value becomes as small as possible. The objec­ a.1 = unit cost of oversupply, tive function in general form, that is, Equation I, cannot be used in ~J = unit cost of undersupply, and practical situations. A special objective function for this research 81 = x/81. problem was developed as follows: Since the PTB system consists of several service locations for which resources should be allocated, the total expected oversupply Minimize Li= 1 c1 (x) and undersupply cost for the whole system would be as follows: Subject to: Li= 1 x1 ~ N, X1>0,j= 1,2,3 ... ,n (2) where (6)

c/x) =expected cost at segment} when x1 is allocated to}, However, if the resources and demand were assumed to be divis­ x1 = resource for allocation, for example, space, ible, then x1 and y are continuous variables that can take any non-neg­ n = total number of service locations inside the PTB, and ative real values. In this case following the same procedure of indi­ N = total amount of available resource. visibility, the total cost function for segment} would be as follows: The main objective in Equation 2 is to minimize the expected cost function. The difference between Equations I and 2 is that in Equa­ tion 2 not all of the resources need to be allocated. The expected cost (7) at each location is a function of allocated resources. As mentioned earlier, there are two types of costs associated with the allocation of resources, that is, oversupply and undersupply cost. Moreover, allo­ where F/y) =cumulative distribution of demand at segment}. cation· of resources depends on the demand placed on the facility. If µ1 is defined as the mean of F1, then by using the principles of As a result, the expected cost is also a function of demand. The probability theory, the preceding equation would finally simplify to: demand at each location is uncertain and depends mainly on the flight schedule. Taking all the variables into consideration, the total (8) expected oversupply and undersupply cost for the PTB is found as follows: Assume that y is the demand variable at each service location and Considering that 81 = x/81, then the derivative of the preceding p/y) is the probability mass function for variable y at segment}. equation with respect to x is as follows: This means that the probability of having y units of demand at seg­ 1 ment} is p/y). It is also assumed that each unit of demand needs 8 1 (9) units of resource at segment}, for example, the amount of space that each passenger occupies. If x1 is the resource allocated to segment}, From Equation 8 and its derivative (F1 is increasing) it is clear that and a.1 is assumed to be the unit cost of oversupply at location}, then the function is convex with respect to variable x1, which means that the oversupply cost at this location is as follows: there is a minimum point in the function (Figure 2). According to Figure 2, the operation cost is high before reaching (3) its optimum point. This is referred to as undersupply cost, which is the cost of physical and psychological discomfort perceived by pas­ where 8 = int(x/8). 1 sengers because of the lack of adequate resources. The real value of If the resources allocated to segment j were less than required, these costs to the airport operator is difficult to estimate. In some then there would be an undersupply cost. Following the same cases they may be roughly· approximated by the frequency of com- Parizi and Braaksma 37

Min. Cost

Optimumxj Resources

FIGURE 2 Cost function of PTB segments with respect to resource value.

plaints and critical journalism or the loss of potential customers. ure 3). In the first part of the practical procedure, the PTB is simu­ The operational cost increases right after the optimum point. This is lated to perform as a real terminal. The simulation model will be the oversupply cost, which is the cost of providing resources beyond discussed in the next section. The simulation is run for a number of what is required. Having information about operational and main­ days or weeks, and the population statistics are collected for each tenance expenses, this unit cost is possible to estimate. However, segment of the PTB during the whole running period. These statis­ the objective function would be a series of nonlinear separable con­ tics are analyzed to arrive at the probability mass function (PMF) of vex functions that have to be optimized. If the values of OLj, J3j and demand for each PTB segment. To be as realistic as possible, the the demand function were known, then there would be some ana­ operating day is divided into short time periods, for example, 1 lytical approaches to solve such problems as Equation 1, in which hour, and the PMF for each time period is obtained. The PMFs and the total amount ofresources would be allocated (6). The objective the values of OL and J3 are used as input to the second part of process, function of this research problem is more complex than the con­ that is, the optimization algorithm. The optimization algorithm ventional ones because of the fact that the sum of allocated determines the optimum value of required resources for each seg­ resources could be less than or equal to the maximum resource ment at each instant of time. The output of optimization algorithm available. Moreover, due to the stochastic nature of passenger is a variable time-resource diagram. arrivals and departures at the PTB, no specific mathematical func­ The sum of optimum resource values from all segments multi­ tion can represent the actual demand on the system at each instant plied by the unit cost of providing resources is the optimum cost of of time. Therefore, demand function may be found either through operating the PTB at each instant of time. If all the conditions are an exhaustive data collection exercise for a long period of time, or met, the operational and maintenance cost will be a function of by using a simulation approach. demand distribution. However, in the third part, it is recommended that the results of the optimization procedure be implemented on the site by some sort of flow management and control technique. The PRACTICAL OPTIMIZATION PROCEDURE existing resources and the traffic flow should be managed in such a way as to provide a diagram as close to the optimum time-resource The ideal process was believed to be difficult to apply in a real ter­ plan as possible. The details of the optimization algorithm and the minal for the time being. Therefore, the three parts of the ideal flow management and control model are ongoing research. The process were modified to develop a more practical procedure (Fig- results will be the subject of our next paper. 38 TRANSPORTATION RESEARCH RECORD 1506

Forecast Traffic, Range Of Nominal Schedules

Simulation Operational Model ofPTB Guidelines

Demand Function of All PTB Segments

Levels of Service, 0 ~-----tM~-----1 Unit Costs, a., ~

Optimization Model

- Optimum Resource Values - Variable Time-Space Plan

Associated Costs I ------:-----

•------1 Flow Management & Control Techni ues

Resources Allocation Based On the System Constraints

FIGURE 3 Practical procedure of optimum PTB resource utilization.

PTB Simulation Model Recently, because of new developments in software technology, there has been a lot of interest to use object-oriented programming Airport terminal simulation models have been developed for more (OOP) in PTB simulation. The object-oriented approach has some than 30 years (7). Through the literature review on existing and cur­ advantages over conventional languages used in simulation. In sum­ rently used simulation models, it was found that models currently mary, the OOP approach is capable of providing immense pro­ available do not respond to the types of issues associated with PTB gramming flexibility, greater reduction of input requirements, ease operation and management. They require extensive programming to of operation, and user friendliness (10,11). OOP also provides fully fit with any specific configuration of a terminal, take a long time to interactive execution with a high degree of animation and graphics process any particular run, and require too much detailed informa­ capabilities. The models in the OOP concept are built in terms of tion for every run (8). An additional problem with currently used real world components of the system, as opposed to reducing com­ simulation models is occasional failures to address important ponents to a series of mathematical relationships and writing com­ aspects of terminal operations (9). puter programs to invoke those relationships. In spite of all of its Pariz,i and Braaksma 39 advantages and interest in the airport industry, a PTB simulation throughout the model. The structures are declared as "unshared," model using OOP is not yet publicly available. meaning that each passenger will preserve its own copy of data, However, for this research, a comprehensive search was under­ including flight number, gate, departure time, and so forth, while taken to find the most recent and state-of-the-art simulation tool. A passing through different segments of the PTB. Several functions new object-oriented graphical modelling environment, the for transaction routing and passenger arrival time sorting are SES/workbench from Scientific and Engineering Software Incor­ declared in the functions_decs node. In the param_declaration node, porated, was found and used to simulate PTB operation (J 2). The several input variables are declared as parameters to include para­ SES/workbench is an integrated collection of software tools for meters in the model. Parameters can be changed during the run­ simulation and evaluation of complex systems, such as computers, time, which makes the model user friendly. large software systems, data communication networks, and micro­ The generate_planes submodel is the planeload-generator sub­ processors. The SES/workbench consists primarily of SES/design, model. A transaction, "seed," is generated, reads each line of the SES/sim, and SES/scope. The graphical representation of the sys­ schedule, and generates another transaction called "plane." The tem is built by using the design interface module, SES/design, in plane transaction is routed to either generate_arrive_pax or gener­ which objects are created to represent the various components of the ate_depart_pax according to its sector, that is, arrival or departure. system. The graphical representation is converted to an executable The generate_arrive_pax submode} generates arriving passen­ simulation model by SES/sim, a translation and simulation module gers. The plane transaction enters the· submode} and waits in the based on the C and C + + programming languages. Finally delay node for its event time. When the transaction reaches its event SES/scope is an animation module that provides the ability to time, it generates the number of passengers according to a normal observe and debug an executing simulation model. probability distribution, with the mean of average deplaning rate In summary, an SES/workbench model is composed of one or defined by the user. The accumulated number of passengers gener­ more submodels, each represented by an extended directed graph. ated for each increment is compared with the total number of pas­ The basic components of a graph are nodes, arcs, transactions, and sengers of each aircraft. Once the array of passenger transactions resources. Transactions are entities that flow from node to node is generated, the plane transaction sends them to the submode! along the arcs. Each transaction represents a process to be executed. process_arrive_pax. Each transaction may carry with it an arbitrary user-defined data The submodel generate_depart_pax generates the enplaning pas­ structure. Each node in a model represents the manipulation, for sengers and sends them for processing. The plane transaction enters example, allocation, or release of a physical or logical resource, or the submodel and after reaching its event time generates the time some other processing step in a transaction's life. Each arc connects that each departing passenger arrives at the airport according to a two nodes and is directed from one node to the other. It represents triangular probability distribution. In the triangular distribution, the a path along which a transaction may flow from one node to another. minimum is defined when the first passenger of the flight arrives at Each resource represents some physical or logical component for the PTB, the maximum as the time when the last passenger of the which transactions compete. flight arrives at the PTB, and mode as the time when the maximum number of passenger arrive at the PTB. The submodel process_arrive_pax models the activities of Simulation Model Framework deplaning passengers and meeters. Passengers unloaded from the aircraft go to the submodel concourse, to be explained later. Pas­ The PTB simulation model (PTBSIM) is designed to predict the sengers coming out of the submode} concourse will be routed movement of passengers, greeters, and well-wishers for a given ter­ . according to their region, that is, domestic, international, and so minal design and a candidate commercial ·aircraft schedule. forth. On arrival of domestic passengers to the baggage claim area, Throughout the model development every effort was made to keep meeters will be generated according to a uniform probability distri­ the model as simple and user friendly as possible. Although the bution ranging from 0 to 2. The international passengers will go model was developed based on a given terminal design, it is very through the preliminary inspection lines (PIL) (Canada Customs easy to adjust the model for any type of PTB due to its object­ and Immigrations), secondary customs, immigration, the baggage oriented aspect. One can easily change, delete, or copy any node, claim area, and the arrival lobby. The service times for all these arc, or submodel to get the desired design. The only required input activities are drawn from some probability distributions in which to the model is an aircraft schedule, which can be entered using any mean and standard deviation are based on historical data (1,13,14). text editor. The aircraft schedule is used to generate passengers and The process_depart_pax submodel models the behavior of nonpassengers entering the PTB. The operating day is divided into enplaning passengers and well-wishers. Departure passenger trans­ equal time increments, for example, 1 minute long. It should be actions are accompanied with the well-wishers. Well-wishers are noted that the size of increments can be any value, for example, generated from an integer uniform probability function. Almost all from milliseconds to hours. the passengers and well-wishers go through the ticket lobby. A per­ The PTBSIM was developed as one main module, consisting of centage of passengers either are preticketed or go to the express six submodels, that is: generate_planes, generate_arrive_pax, gen­ · check. Each major airline and its allied carriers is represented by a erate_depart_pax, process_arrive_pax, process_depart_pax, and service node with some number of servers. Passengers are directed concourse. In addition, several functions were developed to define to the service nodes according to their flight numbers. Depending the workload and transaction routing. These functions and some on how much time is left for each passenger before departure, the parameters are declared in the three declaration nodes. A brief passengers may stay in the waiting and concession area. If the time description of declaration nodes and submodels follows. left for the passengers is too short, then the passengers will experi­ In the main_declaration node, two structures are declared as ence only the walking-time delay. Passengers and their companions plane and passenger. These two structures consist of a collection of proceed to the security booths and gates through corridors or some variables (information) which each plane or passenger should carry vertical transportation facilities, such as escalators, elevators, or .. llD!l l!lil IR ~ Module: ottawa ~ 1111 ... ie. 1111!!11!!111111111!11 Submode!: • . . ------~------·--·- -- i.e., domestic or international.

int

staff_manager pil '.C§J;' custom boot~'Rf Population :O Power :1 Servers :3 ------~------.0.1 begio_arri~~llo-,,,_~,~er

I Reference end arrival fto coocoodL~l~j_____,.,...___o~."~'~l ______l---+l•,~- ..... -a~ival lobbyl B

dom meeters dom_baggage_claim I

Scope Set ings

> Traffic Con rol {3868,greeters,O,NULL} @ 500.049 traversed arc from staff manager to ticket lobby in submode! {3868,greeters,O,NULL} @ 500.049 entering process_depart_pax.ticket_lobby (pop=41) {3868,greeters,O,NULL} @ 500.049 at process_depart_pax.ticket_lobby scheduled to be activated at time=503.164 {3868,greeters,O,NULL} @ 500.049 deactivated at process_depart_pax.ticket_lobby (3869,greeters,O,NULL} @ 500.049 activated at process_depart_pax.generate_greeters {3869,greeters,O,NULL} @ 500.049 traversed arc from generate_greeters to staff_manager in submode! process_depart_pax {3869,greeters,0,NULL} @ 500.049 entering process_depart_pax.staff_manager (pop=l) {3869,greeters,O,NULL} @ 500.049 traversed arc from staff_manager to ticket_lobby in submode! process depart_pax {3869,greeters,O,NULL} @ 500.049 entering process_depart_pax.ticket_lobby (pop=42) {3869,greeters,O,NULL} @ 500.049 at process_depart_pax.ticket_lobby scheduled to be activated at time=501.4 ,:.'{ {3869,greeters,O,NULL} @ 500.049 deactivated at process_depart_pax.ticket_lobby ;;.!·f: {3849,meeters,O,NULL} @ 500.061 activated at process arrive pax.dom baggage claim ~~ {3849,meeters,O,NULL} @ 500.061 exiting process arri;e pax.~om baggige clai; (pop=121) '(,, {3849,meeters,O,NULL} @ 500.061 traversed arc from dom=baggage=claim t~ di 157 in submode! process arrive_pax ),< <1019> step {port} submode!; :Q~ {3849,meeters,O,NULL} @ 500.061 entering process_arrive_pax.di_157 (pop=ll J\':\ {3849,meeters,O,NULL} @ 500.061 traversed arc from di 157 to arrival lobby in submode! process_arrive_pax ·{@:=::~ ~:::•-C-o_mrn_a_n_d_)------•

FIGURE 4 Illustration of SES/design and SES/scope window. Parizi and Braaksma 41

moving belts. Only passengers are allowed to pass the security and ification form may be checked for tracing the problem. After the to go to the concourse. It should be noted that the number of servers debugging process, one may calibrate the model using the anima­ in each service node, that is, PTB personnel, is dynamically man­ tion capability. SES/scope provides a detailed animation of the aged over time by a set node called staff_manager. model events as the model runs. Modellers also have control over The concourse submode] is a waiting area with several gates. Both the events that they choose to see animated. Some of the events that arrival and departure passengers will go through the concourse area. can be animated are transaction movement, transaction tracking, The arrival passengers experience a delay time equal to their walking transaction creation and destruction, service and delay nodes, and distance divided by their walking speed, and the departure passenger queue entry and exit. enters the concourse and waits until the final boarding call. Therefore, For example, as a passenger (transaction) flows through the PTB, each passenger will experience a different amount of delay time. The for example, traverses arcs and enters nodes, it is represented by a reason for making the model into six submodels is to make the model rectangle containing a "T" (for transaction) followed by the trans­ flexible enough for possible adjustments of specific PTBs. action identification number (Figure 4). The most recently traversed arc is thickened considerably. This permits one to observe visually the paths that specific transactions follow as they flow through the PTBSIM Evaluation model. Modellers control the scope of animation by using the but­ tons and defining them according to their own needs. For example, When the basic layout of the model was constructed, the SES/scope one may choose to animate a specific category of passengers, a spe­ was used to calibrate the model. SES/scope allows the modeller to cific service node, or a submode!. Using the SES/scope, the debug­ interact with, control, and debug the model while it is running. It ging process was done and the PTBSIM was finetuned based on also allows one to watch an animated display of the model's exe­ data obtained from the Macdonald-Cartier International Airport in cution and to debug the model should it behave in unexpected ways. Ottawa, Canada. One can interact with the animation of a running model through the The general objective of the validation procedure for PTBSIM SES/scope window below the SES/design window. An example of was to demonstrate the extent of agreement between model outputs the window is indicated in Figure 4. All the trace messages and and corresponding data obtained at the airport. Data observed for other information about the model's state are displayed in the this purpose are time series of flow and queue length at passenger SES/scope window. One can enter control commands at the com­ processing facilities. Data were collected by stationing observers at mand line prompt below the SES/scope window. At the top of several locations throughout the PTB for simultaneous observation SES/scope window is a banner c_ontaining several buttons that are of the population at each processing unit. The model is also capable used to control various animation parameters, and fields that display of producing time series data for direct comparison with field obser­ information about the current state of the model. vation. The outputs from the PTBSIM and observed data versus While SES/scope is active, one may examine specification forms, time were plotted on the same pair of axes for visual comparison. defined for the nodes in the graph, displayed in the SES/design win­ The model provided generally good representation of those facil­ dow. In the case of any unexpected behavior of the model, the spec- ities surveyed, as illustrated in Figure 5. More complete discussion

160 Observed 140 ...... Simulated 120 c: 0 ! "+:l ~ "3 100 ~ Q.. ~ 80

60 \====. II 40 ...... ,•' ~J 20 v 0 860 880 900 920 940 960 980 1000 1020 Time of day, (min)

FIGURE 5 Comparison of simulation outputs and observed data at baggage claim area. ottawa rocess depart

•~~~~~~~~~~~~~~~~~~~~~~~~~~~-~~~~~- ,­ - - - -)- - - - total_enplaning_ time - - , I I

128 256

e

wait_n_concessions Reference end_depart to concourse

ttings

Control ... { 1537, greeters, O,NULL) 393.137 traversed arc from generate_greeters to staff_manager in

,_, .. { 1537, greeters, 0 1 NULL) 393 .137 entering process_depart_pax. staff_manager (pop=l) :~~,:. (1537,greeters,O,NULL) 393.137 traversed arc from staff_manager to ticket_lobby in submode! process_depart_pax K11537,greeters,O,NULL} 393.137 entering process depart pax.ticket lobby (pop=24)

.~ { 1537 1 greeters, 0 1 NULL} 393.137 at process_depart_pax. tlcket_lobby -scheduled to be activated at time=396- 971

''·'• { 1537 1 greeters, O, NULL} 393. 137 deactivated at process_depart_pax. ticket_lobby {1482,departure_pass,O,NULLJ @ 393.183 activated at process_depart_pax.aca_and_allied (; {1482,departure_pass,O,NULL} @ 393.183 de-q'd at process_depart_pax.aca_and_allied {1482,departure_pass,O,NULL} @ 393.183 exiting process_depart_pax.aca_and_allied (pop=9) {1482,departure_pass,O,NULL) @ 393.183 traversed arc from aca_and_allied to di_l69 in submode! process_depart_pax {1482,departure_pass,O,NULL) @ 393.183 entering process_depart_pax.di_169 (pop=l) {1482,departure_pass,O,NULL) @ 393.183 traversed arc from di_l69 to wait_n_concessions in submode! process_depart_pax

{1482 1 departure_pass 1 0 1 NULL} @ 393.183 entering process_depart_pax.wait_n_concessions (pop=78)

Command>

FIGURE 6 Real-time graphic illustration of statistical results.MDNM/. Parizi and Braaksma 43 on the results of simulation can be found elsewhere (15). The model The use of SES/workbench to simulate the PTB instead of using can also be validated in some degree through SES/scope. Running more conventional simulation languages reduced the simulation in SES/scope, one can see the graphic illustration of some statistics, time substantially without reducing the overall accuracy of results. such as population, queue length, and so forth. Figure 6 indicates However, to implement the practical optimization procedure, the population statistics against time for several segments of the more research is needed. At present, research is proceeding to PTB. The modeller would be able to change the parameters and develop a heuristic optimization algorithm that can be combined observe the consequences graphically. The graphs can be zoomed with the simulation model. The procedure will be tested on differ­ for more detail illustration. ent passenger terminals and the results will be compared. Although the visual approach indicates an agreement, the extent to which the model can replicate the existing situations is another important aspect to statisticians. PTBSIM also outputs a list of sta­ ACKNOWLEDGMENT tisticsc for example, mean, standard deviation, maximum, and min­ imum for some parameters defined throughout the model, such as Financial support for this research was derived from Grant No. population, queue length, waiting time, and utilization. These sta­ 04P8927, National Research Council of Canada. tistics are very useful for overall performance evaluations of the model. In addition, the Statistical Package for Social Science (SPSS), a program utilizing the least square method, was used to REFERENCES find the degree of correlations between the observed and simulated values (16). By looking at the results and taking into account 1. Special Report 215: Measuring Airport landside Capacity. TRB, the stochastic nature of passenger activities, it appears that the National Research Council, Washington, D.C., 1987. model can reasonably predict the behavior of passengers in the 2. De Neufville, Richard, and Michel Grillot. Design of Pedestrian Space In Airport Terminals. Journal of Transportation Engineering, ASCE, PTB. As already mentioned, by changing some input variables, any TEI, Jan. 1982. terminal building can be modeled. Moreover, in a specific PTB, any 3. Braaksma, J.P. A Computerized Design Method for Preliminary Airport operational plan can be generated and tested very easily. Therefore, Terminal Space Planning. Technical report, Department of Civil Engi­ the PTBSIM is not only a simulation tool but an evaluation tool neering, University of Waterloo, Ontario, Canada, 1973. 4. Ashford, Norman. Airport Operations. John Wiley & Sons, Inc., 1984. as well. 5. Lerner, A.C. Characterizing and Measuring Peiformance for Ai1port Passenger Terminals. Transportation System Center, U.S. Department of Transportation, July 1990. 6. Ibaraki, Toshihide. Resource Allocation Problems. MIT Press, Cam­ Conclusions and Future Research bridge, Mass., 1988. 7. Mumayiz, S. An Overview of Airport Terminal Simulation Models. The optimization procedure described in this paper can be used as Presented at 69th Annual Meeting of Transportation Research Board, a utilization plan to operate the PTB at its minimum total oversup­ Washington, D.C., Jan. 1990. 8. McKelvey, F.X. A Review ofAirport Terminal System Simulation Mod~ ply and undersupply cost. The procedure can also be used as a short­ els. Transportation System Center, U.S. Department of Transportation, or long-term planning tool. Given a nominal aircraft schedule, the Nov. 1989. planner would be able to simulate the basic activities and services 9. Odoni, A.R., and R. Neufville. Passenger Terminal Design and Com­ required for a passenger terminal. Using the statistics obtained from puter Models. Paper prepared for FAA, U.S. Department of Trans­ the simulation as input to the optimization model, a modeller will portation, July 1990. 10. Smith, N.D. Concepts of Object Oriented Programming. McGraw-Hill, get a variable time-resource diagram. The diagram could be daily, Inc., New York, 1991. weekly, monthly, or yearly depending on the accuracy of the analy­ 11. Mumayiz, S.A. Interactive Airport Landside Simulation: An Object sis. The designer can use these diagrams to prepare a more flexible Oriented Approach. Presented at 70th Annual Meeting of Transporta­ and efficient physical layout. Therefore, there would be a better tion Research Board, Washington, D.C., Jan. 1991. association between the physical and operational plans at the early 12. SES/workbench: Users Manual. Scientific and Engineering Software, Inc., Austin, Tex., March 1992. stages of planning. 13. Interim level of Service Standards. Canadian Airport Systems Evalua­ The idea of flow management and control can respond to some tion, Transport Canada, 1977. real-time events which may happen due to an uncertain economy, 14. Airport Terminal Reference Manual. International Transport Associa­ bankruptcy or replacement of a carrier, traffic demand changes, or tion, Geneva, Jan. 1989. 15. Mahmoud, Parizi S. A Practical Procedure for the Optimum Operation even natural factors such as inclement weather. of Airport Passenger Terminal Buildings. Ph.D research proposal. Using the PTBSIM, the airport operator would be able to place Department of Civil and Environmental Engineering, Carleton Univer- - the demand on the PTB system and observe its operation. The oper­ sity, Ottawa, Canada, Feb. 1994. ator can also interactively test the PTB operation under different 16. Nie, Norman H. Statistical Package for the Social Sciences, 2nd ed. load conditions or operational plans. PTBSIM can also be used as McGraw-Hill Book Company, New York, 1985. an operations tool for the operating staff to help them to maintain a Publication ofthis report sponsored by Committee on Airport Terminals and reasonable level of service through the PTB. Ground Access. 44 TRANSPORTATION RESEARCH RECORD 1506

Analysis of Moving Walkway Use in Airport Terminal Corridors

SETH YOUNG

This paper explores the use of pedestrian conveyor systems, otherwise those passengers carrying baggage, airports, in theory, are ideal known as moving walkways, in long public corridors such as those locations for moving walkways. found in major commercial airports. The investigation includes a brief Airport terminals have had a unique experience with passenger comparison of moving walkways with other primary modes of airport conveyor systems. Over the years, airports have acted as testing terminal passenger transportation and an empirical study of the use of moving walkways through analysis of passenger conveyors at the grounds for technological modifications to passenger conveyors. In United Airlines Terminal at San Francisco International Airport. The addition, airports have provided unique arenas for competition empirical study investigates the physical characteristics of several con­ between passenger conveyors and other passenger-mobility sys­ veyors and their locations within the airport terminal. The study also tems, such as electric courtesy carts, people movers, and buses. examines the passengers that traverse the corridors where the moving With each mode's inherent advantages and disadvantages, present walkways are located. Characteristics of the passengers, along with conveyor technology has found its niche within the airport terminal their "mode choice" of transport along the corridor were recorded. With these data, a brief examination of current passenger use is made, with environment. This paper will discuss briefly the characteristics that an emphasis on how travel speeds vary with each mode. In addition, have determined its current limited success and provide insight into implications are drawn concerning a passenger's mode choice, by how modifications to the passenger conveyor may result in success means of two discrete choice Logit models. The paper briefly compares in more areas of society. the findings from the empirical analysis with similar studies performed in Europe in the 1970s. The comparison determines improvements that have been made since the European studies. Finally, the paper draws PRESENT DAY PASSENGER MOBILITY SYSTEMS some speculations as to how characteristics of passenger conveyors may be altered, in hopes of improving their services and ultimately Airports are ideal locations for pedestrian mobility systems for a increasing their niche in the pedestrian transport market. variety of reasons. The grand scale of most airport terminals and the need for baggage-encumbered passengers to move long distances First proposed over 100 years ago, the moving walkway, or motor­ quickly creates a demand for enhanced mobility. ized passenger conveyor, has been considered an innovative mode Airline hub-and-spoke route configurations have increased terminal of pedestrian transportation. The first public operational moving sprawl while reducing a passenger's time frame for making connect­ platforms carrying pedestrians were found at entertainment com­ ing flights. Therefore, interchanging passengers must be able to move plexes (such as the 1893 World's Colombian fair in Chicago) and through the terminal quickly to switch flights. Furthermore, passen­ were considered as novelty items. The first effort to implement a gers must deal with the long distances associated with large-scale ter­ conveyor system solely for the purpose of serious passenger trans­ minal buildings that are continually growing as air traffic grows. port occurred in 1904, with a proposal to build a continuous mov­ There are a number of pedestrian movement technologies available ing walkway subway under 34th street in Manhattan, New York, to airports. Presumably, the primary reason for such technologies is but was never implemented successfully. Few such systems were to reduce the passengers' travel times throughout the terminal envi­ actually operational in the United States before 1950, the most suc­ ronment. The four primary technologies in use today are as follows: cessful of them being Cleveland, Ohio's "Rolling Road," which transported pedestrians as well as horse and carriages from the low­ • Courtesy carts, lying warehouse district to the downtown, some 20 meters higher in • Buses, elevation. Virtually all operational moving walkway systems were • Automated people movers (APMs), and - defuµct by the early 1950s. • Moving walkways. Modern times have shown a rebirth in the passenger conveyor. The moving walkways, however, have been relegated to particular Articles by Leder (J), Smith (2), and Sproule (3) present compre­ market niches. Conveyors are now primarily found at indoor facil­ hensive reviews of each of the above modes, describing the tech­ ities such as sporting arenas and auditoriums. Most significantly, nology of each mode, performance measures, and inherent advan­ they are found in major transportation-oriented facilities, such as tages and disadvantages. Table 1 compares the four primary rail stations, parking garages, and airports. Airports, in fact, are the transport modes. The advantages, disadvantages, and primary mar­ most predominant users of moving walkways. Because of the large ket niches for each mode are described. areas required by aircraft for maneuvering, the large number of pedestrians passing through airports, and the large percentage of Courtesy Carts

Institute of Transportation Studies, University of California, Berkeley, Courtesy carts are highly maneuverable, electric powered, rubber Berkeley, Calif. 94720 tired vehicles that can navigate though a terminal concourse shared Young 45

TABLE 1 Performance Characteristics of Passenger Mobility Systems

Mode Typical Operating Speed Headway Capacity

Courtesy Cart 4 - 8 km/hr. variable 5 pax/cart 150-200 pax/hr Bus 16 - 55 km/hr. 5 - 15 min. 15 - 60 pax/bus 500 - 1500 pax/hr APM 13 - 80 km/hr. 1 - 5 min. 1,000 - 14,000 pax/hr. Moving Walkway 30 m/min. none typical: 5,000 pax/hr.

with pedestrians, furniture, fixtures, and building components with APMs typically have high passenger acceptance because of their ease. Carts are typically used in three cases. outstanding safety and service record. However, these systems have high facility and maintenance requirements. As a result, APMs are • To transport mobility-impaired passengers who cannot walk best suited to relatively high rider levels over routes longer than 300 long distances or whose walking speeds are well below normal; meters, although shorter alignments in specialized situations do exist. • To transport passengers making close connections when above normal walking performance would be insufficient; and • To provide organized service over a fixed route. Moving Walkways/Passenger Conveyors

The flexibility of deployment and scheduling, along with the The conventional moving walkway is a pedestrian-carrying device potential for operation without dedicated building infrastructure are on which passengers may stand or walk. Propulsion is provided by two primary advantages of the carts. The largest disadvantages a treadway that moves at a constant, uninterrupted speed and offers include unscheduled, hence unreliable, service, the low capacity of point-to-point service. Nominal lengths vary from 30 to 120 m. the mode, and the potential of corridor gridlock in congested areas. Local building codes often govern maximum lengths on the basis of Furthermore, courtesy carts are generally disliked by those pedes­ emergency exit requirements. Treadway widths typically range trians who do not use the mode, due to their apparent intrusion into from 100 cm (most prevalent) to 140 cm. Inclines of up to 15 the pedestrian corridor, and the potential for pedestrian safety com­ degrees are possible. Treadway speeds are typically between 25 and promises. Courtesy carts can serve an important role in assisting the 35 m/min. 30 m/min is the typical operating speed. Regard for pas­ mobility-impaired and connecting passengers with a time shortage, senger safety prevents higher operating speeds. The capacity of a but they are not viable for significant ridership levels because of moving walkway varies with its environment, depending on the their physical design and operating environment. speed of the tread way and the foot speed of passengers, among other variables. Practically, moving walkway systems have been found to have capacities of about 5,000 passengers per hour per direction. Buses Walkway facilities have their inherent advantages and disadvan­ tages. A moving walkway speed of 30 m/min is considerably lower Buses are rubber-tired, driver-steered vehicles operating mostly on than the average pedestrian walking speed of 70 m/min. Another streets and roads in mixed traffic. At airports they typically operate disadvantage is the small width of the typical walkway. This is a on terminal frontage and circulation roadways on a nonexclusive particular problem in airports where luggage-laden passengers basis, providing both scheduled and on-demand service to defined walking on the conveyor wish to pass other luggage-laden passen­ curbside stops that are easily relocated. They are used typically for gers standing still on the conveyor. Further disadvantages of the transporting passengers between major airport facilities, such as system are the barriers to cross-concourse traffic, the inaccessibil­ between terminal buildings, parking areas, and regional public tran­ ity to wheelchair or otherwise mobility-impaired passengers, and sit systems. the inflexibility of the system. Some advantages of the system Some advantages of this mode are its flexibility, relatively low include the fact that there are no headway and, hence, no waiting cost, and high capacity. The biggest disadvantage is its observed time for service, unless the arrival of passengers exceeds system quality of service. Buses are often considered "uncomfortable," and capacity. The system is perceived to be safe and simple and may be air passengers with baggage often are unwilling to tolerate either integrated easily into any airport terminal environment. Mainte­ crowds or long wait times. Because of their curbside stops, buses nance of the system is also relatively simple. Careful maintenance are inconvenient for connecting passenger transportation, and air­ planning is required, though, because any system stoppage during port congestion keeps service speeds low. periods of terminal activity could cause severe inconvenience and because there is no quick-fix or backup system typically available. An excellent review of the historical evolution of moving walk­ Automated People Movers way technology, including a description of the three main types of moving walkways in use today (rubber belt systems, cleated pallet An APM is a class of public transit characterized by its automatic systems, and rubber covered pallet systems), may be found in John driverless control of discrete vehicles operating on exclusive rights­ Tough and Coleman O'Flaherty's book entitled Passenger Con­ of-way, using a specialized guideway to control the vehicles' path. veyor: An Innovatory Form of Communal Transport (4). Because APMs are proprietary systems, many technological fea­ Knowing the inherent characteristics of moving walkways, such tures vary between suppliers. pedestrian transport systems appear in theory to fit quite well in the 46 TRANSPORTATION RESEARCH RECORD 1506 airport environment. Despite this theory, however, little has been • Whether the passenger was a business- or leisure-type traveler; studied concerning the use of the conveyors empirically. What fol­ and lows is one such study. Specifically, a case study of the moving • The number of bags carried by the passenger. walkways at the United Airlines Terminal at San Francisco Inter­ national Airport is made, with the goal of determining who uses the The above characteristics were made using the best judgment of moving walkways, how, and when, depending on the characteris­ the observers. Although age and sex characteristics appear clear to tics of the moving walkways, the corridors in which they are assess, passenger type may be unclear. Characteristics of the per­ located, and the passengers that travel the terminal corridors. son's attire and type of baggage were the two factors most keenly observed to determine the purpose of the passenger's trip. For instance, an adult in a business suit carrying a briefcase was noted ANALYSIS OF MOVING WALKWAYS: to be a business traveler, whereas a child in casual clothes carrying UNITED AIRLINES TERMINAL, SFO a teddy bear was noted as a leisure traveler. The amount of baggage was recorded in the following manner. Methodology For the purposes of this study, any item larger than a purse was con­ sidered to be a baggage item. No discrimination as to the weight of The United Airlines terminal at the San Francisco International Air­ an item was made. Although the method an item was carried (e.g., port has four sets of passenger conveyors for public use. They are over the shoulder, toted along wheels, or lifted) was recorded, located as follows: analysis later revealed that this characteristic had little effect on conveyor use and was dropped from the study. • Between gates 80 and 84, after security check; As the traveler crossed the conveyor threshold, it was recorded • Immediately after security check, before gates; whether (s)he entered onto the conveyor or bypassed it. If the • Corridor between parking garage and check-in counters; and person chose to use the conveyor, it was recorded whether the • Between gates 84 and 87, after security check. person stood still and traveled at the speed of the conveyor, or walked along the belt. Finally, the duration of travel from the Conveyor banks I through 3 were visited on Sunday, April 24, entrance threshold to the exit threshold was recorded. The exit 1994, between 3:00 p.m. and 6:00 p.m. The following physical threshold was defined as an imaginary line where the conveyor belt characteristics of each conveyor system were surveyed: ends, marking the point where the presence of the conveyor has no bearing on the passenger's travel. This observation defined the three • Treadway belt speed; modes of transport along the corridor, STAND, WALK, or • The number of belts in each direction; BYPASS, and provided complete measurements on the time each • "Departure" direction = heading toward the farthest gates; passenger needed to traverse the corridor, given his/her mode • "Arrival direction" = heading toward the parking garage; choice. A total of 269 observations were made during the observa­ • The incline of the corridor (in degrees); tion period. • The length of the conveyor; and • The width of the conveyor belt. Analysis of Observed Data Table 2 describes the observed characteristics of the conveyors. At the time of the study, the arrival direction conveyors in con­ Initial analysis of the sample revealed that a vast majority of the veyor bank 2 were closed for maintenance. After recording the sample did use the conveyors in some fashion (i.e., either chose above data, observations regarding how each conveyor was utilized mode STAND or WALK). Of these, approximately one-third of the were made. At each site, pedestrians were selected randomly for conveyor users stood still on the conveyor belt. The distribution of observation on passing a predetermined entrance threshold in the mode choice, as well as the average travel speed and travel time of corridor. This threshold was determined to be an imaginary line each group of passengers, is illustrated in Table 3. along the corridor floor perpendicular to the entrance to the con­ The most striking results to come out of this initial analysis was veyor and approximately 5 m before the entrance to the conveyor. that the average travel speed for passengers using the conveyor This location made it possible to observe passengers who were ded­ (STAND/WALK combined) was only marginally higher than for icated to traversing the corridor but had not yet committed to his/her those who chose BYPASS the conveyors. Moreover, the average mode of transport. During this time, the following passenger data travel time to traverse the corridors was almost 7 sec higher for were collected: those using the conveyors than for those bypassing. These initial results prompted two further forms of analysis, one to further • The passenger's approximate age, to the nearest decade; explore changes in passenger foot speed and one to evaluate the • The passenger's sex; mode choice made by passengers.

TABLE 2 Conveyor Characteristics

0 Conveyor Speed (m/sec) #belts (dep.) #belts (arr). Slope( ) Length (m) Width (cm)

1 0.64 0 85 100 2 0.64 2 0 120 100 3 0.64 +2° arr. 80 100 Young 47

TABLE3 Mode Choice Distribution Mode #Sampled % Total Avg. Speed (m/sec) Avg. Time (sec)

STAND 57 21 % 0.67 135 WALK 146 54% 1.70 59.18 USE 203 75% 1.41 80.47 BYPASS 66 25% 1.36 73.20

Analysis of Changing Foot Speeds be used. Such a data file required data that could not be observed directly in the field. Specifically, the travel times and travel speeds for The above initial observation prompted a deeper analysis of changing the two modes that a passenger did not choose when traversing the cor­ walking pace for those using moving walkways. This analysis stud­ ridor could not be observed and recorded. An estimation of these alter­ ies the changing foot speeds of passengers choosing to WALK. The native choice attributes was made by the following methodology. analysis was performed for each conveyor in each operable direction. For each conveyor in the study, the noted conveyor belt speed Estimation of Alternative Choice Attributes and average walking pace for a bypassing passenger were summed together to determine a theoretical travel speed for a passenger who STAND chooses to WALK on the conveyor without changing natural foot For those passengers who did not choose STAND, the travel speed. This theoretical speed was then compared with the average speed and travel time values for the alternative STAND were walking speeds for each location. The results of these calculations merely calculated as follows: are found in Table 4. The results show a decrease in walking speeds ranging from 0.15 to 0.45 m/sec for passengers walking on con­ Travel Speed = Belt Speed veyors. Further analysis, breaking down the sample of passengers Travel Time = Belt Speed * Length of Conveyor into groups according to similar .characteristics, revealed no signif­ icant departure from the overall decrease in walking speed. WALK This consistent decrease in walking speeds may be primarily due For those passengers who did not choose WALK, a linear regres­ to the physical characteristics of the conveyor. The narrow width, sion model was applied. The regression was based on those pas­ rubber-belt footing, and belt speed of the conveyor all contribute to sengers who did choose WALK, and the characteristics of their passengers slowing their step when walking. Furthermore, the environment when the choice was made. The specific characteris­ walking speeds tended to be slow for conveyors having a higher tics included in the regression were as follows: passenger flow, and hence higher degrees of congestion.

X1 = Corridor/Conveyor Length (m). (1 a)

X2 =Belt Speed (m/sec.). (lb)

Mode Choice Analysis X3 = Congestion level of the conveyor ( 1 = congested, 0 = uncongested). (le)

To explore the mode choice made by passengers, analysis was per­ X4 =The incline (slope) of the corridor (degrees). (Id) formed by the evaluation of two discrete choice Logit models, each X5 = The age of the passenger (to the nearest decade). (1 e) with one independent variable. The first model evaluated the three X6 = The sex of the passenger (1 = male, 0 = female). (lf) mode choices (STAND vs. WALK vs. BYPASS) using the inde­ X7 = The "type" of passenger (1 = business, 0 = leisure). (lg) pendent variable travel speed. The second model evaluated the X8 = The number of bags carried by the passenger. ( 1h) same choices using the independent variable, travel time, which takes into consideration the length of the corridor traversed. (Note: The following analysis was performed using data measured The Lo git· analysis was performed using the ALOGIT computer in U.S. units. Resulting formulations were adjusted to metric units software package. To successfully run ALOGIT, a proper data set must after analysis was performed.)

TABLE4 Changing Foot Speeds, by Conveyor

Conveyor 1 Dep. 1 Arr. 2 Dep. 3 Dep. 3 Arr.

Bypass Speed (m./sec.) 1.34 1.36 1.18 1.60 1.40 Belt Speed (m./sec.) 0.64 0.64 0.64 0.73 0.73 Bypass+ Belt Speed (m/sec) 1.98 2.0 1.82 2.33 2.14 Conveyor Length (m.) 85.4 85.4 119.3 79.3 79.3 Est. Travel Time (sec.) 43.08 42.75 65.48 33.97 37.13 Obs. WALK travel time (sec.) 46.97 51.84 88.44 55.00 49.00 WALK+ Belt Speed (m./sec) 1.82 1.65 1.35 1.44 1.62 Obs.WALK foot speed (m/sec) 1.18 1.01 0.71 0.70 0.89 Change in foot speed (ft./sec) -0.16 -0.35 -0.47 -0.89 -0.52 48 TRANSPORTATION RESEARCH RECORD 1506

Table 5 displays the resulting linear regression coefficients that A successful run of ALOGIT using the Travel Speed data pro­ were used to estimate the total travel speed and travel time of the duced the following utility functions for each mode choice: passengers should they have chosen to WALK: WALK: Uw = 0.7241 + 0.076TSw (3a) (Note: All regression equations are of form: STAND: Us = 0.0326 + 0.076TSs (3b) Y= a+ (3,X, + f32X2 + ... + (3,,X,,) (2) BYPASS: U8 = 0 + 0.076TSe (3c)

Initially, only the travel speed independent variable was consid­ A few interesting observations may be made from these functions. ered for regression analysis. Travel time was to be estimated by The most distinguishing characteristic is the positive alternative merely multiplying the estimated speed by the corridor length. Fur­ specific constant in the STAND mode utility function. More impor­ ther consideration, however, brought on the hypothesis that the tantly, the constant is higher than that of the base mode, BYPASS. length of the corridor itself may have a significant effect on the This implies that standing indeed may carry a higher utility for those travel speed of a pedestrian. This may be most prevalent in those whose normal foot speeds are very close to the speed of the belt. who chose to WALK along the conveyors. Along longer corridors, Pedestrians of older ages, as well as those with physical impair­ for example, several passengers were observed to STAND on the ments may easily fit this category (it is interesting to note that age conveyor for a portion of the trip, and then begin walking for the was indeed one of the more significant variables in the travel speed duration. To test this hypothesis, estimated travel time data from the regression analysis). above regression equation was compared with simple time = length Applying the above utility functions into the Logit model reveals * speed results against true data for those who did indeed WALK. some interesting issues. In comparing two hypothetical passengers The regression equation did produce a better match to the true data. with the following travel speed characteristics: The results of the simple formula tended to underestimate those whose true travel time was on the high end of the spectrum. "Healthy" "Impaired" It is interesting to note that belt speed has a negative effect on TSw = 6.0 ft/sec TSw = 4.5 ft/sec (4a, 4b) total travel speed across the corridor. This enriches the above analy­ TSs = 2.1 ft/sec TS = 2.1 ft/sec (4c, 4d) sis of changing foot speeds when walking on conveyors. Another 5 TS = 4.6 ft/sec TS = 3.0 ft/sec (4e, 4f) interesting result of this regression was that the speed increases with 8 8 increasing numbers of bags carried. One explanation for this may the following utility values are derived: be that those passengers with more baggage tended to be in more of a rush to catch their flights than were those with fewer bags. Other Uw = 1.18 Uw = 1.07 (5a, 5b) independent variables appear to have intuitive effects on travel Us= 0.192 Us= 0.19 (5c, 5d) speed and time, which further justifies the use of the equations in Ue = 0.349 Us= 0.023. (5e, 5f) the estimation process.

BYPASS Applying these utility values to a Logit function of the form:

For those passengers who did not choose to BYPASS the con­ P(mode x) = eux + euy + euz (6) veyor, a similar regression analysis was performed to estimate their BYPASS speeds and travel times. The variables used in the regres­ sion were those that would have had an effect on their bypass speed The following mode choice probabilities are found: or travel time, respectively. Table 6 lists the values that represent coefficients in the travel speed and travel time equations. "Healthy" "Impaired" Again, it is interesting to note the increase in travel speed and P(WALK) 0.55 0.57 (6a, 6b) similar reduction in travel time with increasing numbers of bags car­ P(STAND) 0.21 0.23 (6c, 6d) ried. Other variable coefficients appear more intuitive. P(BYPASS) 0.24 0.20 (6e, 6f)

TABLE 5 Regression Coefficients, WALK Alternative

Coefficient Travel Speed (TSw) Travel Time (TT w) ex + 2.485 (t = 1.22) + 54.46 (t = 0.56) ...... ~~ ...... ~.9:99~ ...... (t.~ .. ~9:~9). + o.ot ...... (t.'.".'.o.. qsJ...... ~2 - 1.086 . (t.'.".' .. ~9:79) ...... + 4.79 ...... <~.'.".'.. o ... 2.2.)...... P.3 .. - 0.363 ......

0.07 0.30 Young 49

TABLE 6 Regression Coefficients, BYPASS Alternative

Coefficient Travel Speed (TSe) Travel Time (TTe) a + 1.380 (t = 11.01) -34.00 (t = -3.75) ...... ~! .... 0 ...... :+.. q:3.? ...... (~.:::.P ... ?.~L ...... ~2 .. 0 0 0 0 ...... ~3 ..... ~4 -0.11 ...... (~.:::.. ~q:9.~) ...... :+.. 1.:~~ ...... (t :::.. ~.).q) .... . -~~ + 0.01 ......

....~~-­ . . . . . :'... q:J.3 .... _(t.::: .. 1.-3_7.) ...... ~}:74 ...... <~.:::.. ~q:?7)...... ~? .. + 0.013 ..(t.::: ..~q:~~L - I.68 ...... <~.:::.. ~q:~~L ~8 +0.212 (t=l.41) -3.69 (t=-1.96)

0.12 0.79

These probabilities are highly consistent with the proportion of pas­ Shortcomings to the Above Study sengers' mode choices in the sample data. From the similarities in mode choice probabilities, as well as It is conceded that the above mode choice analysis does have its from direct inspection of the travel speed coefficient of 0.076 in the share of shortcomings. The most prevalent is the fact that the obser­ utility functions, it can be inferred that the sensitivity of travel speed vation process itself led to a series of biases in the sample. to mode choice is quite low. Although in the above comparison the The fact that sampling was only performed on a Sunday after­ probability of choosing STAND does make the shift to exceeding noon in April most certainly resulted in a biased sampling of leisure the probability of BYPASSing, the overall difference in probability passengers. An additional survey performed during a weekday values between the "healthy" person and the "impaired" passenger morning or evening would provide a larger sample of business pas­ are marginal. sengers. It would be prudent to expand the data set to observations An interesting issue arises when comparing the results of the over different periods of a week, and perhaps a year, to collect a travel speed model with the travel time model. An ALOGIT run suc­ comprehensive, unbiased, and perhaps time-sensitive data set. In cessfully made using the above collected data and regression addition, the study may have included several "leisure" passengers derived travel time values resulted in the following utility function who were not passengers at all, but those meeting or seeing off pas­ for the three mode choices: sengers. These people may have behavior patters of their own, which were not recognized in this study and merely were absorbed WALK: Uw = 0.932 + 0.01584ITw (7a) within the leisure passenger category. Furthermore, observer biases

STAND: Us = -1.287 + 0.01584IT5 (7b) of a passenger's age and travel type are in no way insignificant. A

BYPASS: U8 = 0 + 0.01584TS8 (7c) direct response from the passengers themselves would alleviate any prejudices by the observers. The primary issue to strike the observer when comparing travel Some elements of the conveyor environment related to the phys­ time utility functions with the travel speed functions is the magni­ ical environment and to the passenger characteristics were excluded tude of the difference between the variable coefficients and the from the survey. Environmental characteristics such as the presence alternative specific constants. The travel time model has much of windows or other displays, or the presence of destinations (such higher alternative specific constants relative to the independent vari­ as gates, or other corridors) at locations between the start and end able coefficient, marking a significantly less sensitive variable in of the conveyor were considered. Passenger group characteristics travel time. What tends to drive the travel time utility functions are were also excluded. That is, there is no differentiation between a the alternative specific constants, which set an immediate signifi­ passenger traversing the corridor alone with one partner, or with a cant utility ranking. WALK is clearly the most preferable mode, fol­ large group, etc. These characteristics are perhaps significant con­ lowed by BYPASS, and STAND is a distant third choice. Only as tributors to mode choice in the corridor. travel times for STAND increase with rates significantly higher than for BYPASS and WALK will STAND ever become a preferable choice. This is indeed what tends to happen as the length of the cor­ COMPARISON WITH HEATHROW AIRPORT ridors increase, or as conveyors become congested. "TRAVELLA TOR" STUDY It is difficult to decide which of the above Logit models should be used preferably as a basis for any general conclusions about A study conducted by the Loughborough University of Technology moving walkway utilization. However, the utility functions result­ assessed the use of the "travellator" passenger conveyor at Lon­ ing from each analysis suggest that the travel speed model may be don's Heathrow airport terminal 3 in April 1974 (5). The purpose more sensitive to passenger related issues, such as passengers' ages of the study was to determine user behavior on the conveyors, much or number of bags toted. Conversely, the relatively high alternative like this study. Their study, however, focused on safety and com­ specific constants of the travel time model may better describe mode fort issues concerning the conveyors. The findings of the study may choice probabilities derived from the inherent physical characteris­ have resulted in modifications to the design of passenger convey­ tics of the corridor environment itself. ors, including those at San Francisco's airport. 50 TRANSPORTATION RESEARCH RECORD 1506

A preliminary study of user behavior at Heathrow revealed that CONCLUSION a significant proportion (nearly 20 percent) of conveyor users had a non-negligible amount of difficulty with the conveyor. Most diffi­ In the above analysis of pedestrian mode choices throughout the culties related to users losing their balance when boarding or alight­ United Airlines terminal at San Francisco International, we hope to ing the belt. The loss of balance was primarily due to slight move­ provide some insight into who uses passenger conveyors, how, and ments or compulsive jerks experienced when boarding. The results why. By looking at the characteristics of the corridors themselves of the preliminary study led to a more in depth analysis of the travel­ and the passengers who traverse the corridors, discrete choice mod­ lator at Heathrow and a comparison to a similar conveyor in France els based on travel times and travel speed were made. In addition, the (the Montparnasse travellator). phenomenon of passengers changing foot speed was studied briefly. The Heathrow travellator ran for a length of 110 m in a corridor The results of the analysis, although questionable in their statis­ known as the "pier connector" that connects Heathrow's main ter­ tical significance, do provide some insight into the use of passenger­ minal with "terminal 3." The conveyors were 1 m wide and were moving walkways. It is shown that the vast majority of passengers operated at a speed of 0.67 m/sec. Observations of passengers using use the conveyors in some rpanner. Those who use them tend to the conveyor were made by filming pedestrian flows between 6:00 WALK along with the conveyor belt, rather than STAND still. a.m. and 10:00 a.m. This time was chosen to observe peak flows, There are suggestive relationships among passenger characteristics, when large aircraft normally arrive from overseas. Using the film a environmental characteristics, and mode choice. The analysis sug­ total of 290 passengers were studied. Similar to our study, the sex, gests that looking at a travel speed based model is recommended age, and number of bags carried by each passenger were recorded. when analyzing mode choice on the basis of passenger characteris­ Furthermore, the approach speed, step length, and boarding speed tics and that a travel time-based model is preferred when looking at of each passenger were calculated (by counting travel distance per issues of the physical corridor environment itself. frame of film); whether the passenger used the conveyor handrail, The above results lead to the implication that the passenger con­ did a "threshold check" (i.e., adjusted their step length to board the veyor has become a popular mode of transportation not by reducing conveyor), and whether the passenger had any problems boarding passenger travel time, but by acting as a convenience for those pas­ were all noted. sengers who wish to slow their walking pace or stand still while The studies performed at Heathrow and Montparnasse found that traveling the corridor. For this reason, moving walkways have the Heathrow conveyor had considerably more boarding problems found a solid niche in airports for those routes with insufficient (31 percent of sample) than the Montparnasse (20 percent) con­ pedestrian density to warrant other modes, such as APMs, but suf­ veyor. This despite the fact that the Montparnasse conveyor trav­ ficient lengths to preclude walking. Since the Heathrow study, mov­ eled at a higher velocity (0.854 m/sec) than Heathrow's (0.67 ing sidewalks have appeared in more locations and have been. m/sec). improved, and as a result the public seems to be comfortable with The results of the study concluded that the main factor causing their use. However, because moving walkways are perceived conveyor boarding problems was "improper boarding techniques." presently as a convenience rather than a necessity, their full poten­ These techniques mainly involved "over-preparing" to board the tial may not be realized fully. It would be beneficial to passengers conveyor, by altering one's step, grasping for the handrail too early, if moving walkway systems were developed that could capitalize on or looking down when boarding. Their conclusions were supported the low cost and convenience of use without having to pay the price by the fact that the population of users at Montparnasse were currently associated with the mode's shortfalls, such as slow belt younger, business travelers who were familiar with the travellator, speeds, narrow belt widths, and one-destination limitations. Airport whereas the Heathrow users were older, leisure passengers with less terminals would serve as excellent test:.. beds for conveyor improve­ experience on moving walkways. ments in the above areas. Terminals have a continual supply of Whether the Loughborough conclusions are plausible or not, unfamiliar system users. Also, systems easily can be tested in a some of its recommendations for conveyor improvement appear to short-distance configuration before full-scale installation. This ideal have been used in modern conveyor systems, including those in San situation should encourage terminal designers to research any new Francisco International. For instance, the moving handrail was deter­ developments in the moving walkway arena and to consider seri­ mined to be an important aid in maintaining one's balance when ously installing cutting-edge systems that could outperform current boarding the conveyor. Extending the handrail beyond the entrance technology. threshold would help passengers to judge the speed of the system so With the above methods described in this empirical study and the that boarding could be accomplished more successfully. The width above technical considerations, authorities considering the installa­ of the conveyor was determined to affect the ease of use as well. tion or modification of airport corridors with passenger conveyor Conveyors that were too narrow often led to easily obstructed pas­ systems may gain further insight into their potential investments. sageways, leaving less sight and, hence, less preparation for board­ Such insight may also lead to the proliferation of the passenger con­ ing. The study suggested that the width of the conveyor be increased veyor, or moving walkway, further into the realm of pedestrian to at least 1 m. Finally, the addition of instructional signs such as transport. "Keep Walking when Boarding" were suggested. The study suggested that these improvements along with the increased experience the public has with passenger conveyor would REFERENCES reduce the amount of passenger difficulties. During the course of the SFO study, no passengers were observed to have any difficulty with 1. Leder, W. H. Review of Four Alternative Airport Terminal Passenger the conveyors. Although there were no instructional signs evident, Mobility Systems. In Transportation Research Record 1308, TRB, National Research Council, Washington, D.C., 1991. the handrails were extended from the entrance threshold, the con­ 2. Smith, L. L. The Maturing Airport People Mover Field: Four Rounds of veyor was wider than the Heathrow travellator and, probably, the Experience at Tampa International. In Transportation Research Record passengers observed were more familiar with moving walkways. 1308, TRB, National Research Council, Washington, D.C., 1991. Young 51

3. Sproule, W. J. Airport Development with Automated People Mover Sys­ 5. Hawkins, N. M. Behavioural Observations of Passengers Boarding a tems. In Transportation Research Record 1308, TRB, National Research Slow Speed Conveyor. Transport Technology Assessment Group, Council, Washington, D.C., 1991. Loughborough University of Technology, November 1974. 4. Tough, J.M., and C. A. O'Flaherty. Passenger Conveyors: An Innova­ tory Form of Communal Transport. Garden City Press Ltd., Letchworth, Publication of this paper sponsored by Committee on Airport Terminals and Herfordshire, 1971. Ground Access. 52 TRANSPORTATION RESEARCH RECORD 1506

Characterization of Gate Location on Aircraft Runway Landing Roll Prediction and Airport Ground Networks Navigation

XIAOLING Gu, ANTONIO A. TRANI, AND CAOYUAN ZHONG

This study presents an aircraft landing simulation and prediction model. also provide information about the distribution of aircraft landing The model uses simple aircraft kinematics coupled with individual pa­ distance and runway occupancy times for determining optimal run­ rameters to describe the landing process. A multiobjective optimization way exit locations. This procedure is usually carried out in the plan­ and a shortest path algorithm are used to predict the aircraft exit choice ning stage of new runway facilities. and taxiway path in the runway taxiway network. By recognizing pilot motivation during the landing process, several influence factors such as terminal location, runway, and weather conditions are considered in the aircraft landing simulation. Random variables such as aircraft runway BACKGROUND crossing height, flight path angle, approach speed, deceleration rate, and runway exit speed are generated to represent the stochastic landing Earliest efforts to describe aircraft runway landing process are behavior of aircraft by using a Monte Carlo sampling technique. With found in runway exit optimization and capacity analysis (J-3). In real-time input data, the model could provide information on aircraft exit choice, runway occupancy times, and shortest taxiway path to an 1974, Joline (3) used an aircraft deceleration model to predict the assigned terminal location for both the pilot and the air traffic controller runway occupancy time in the runway exit location problem. Based in a ground traffic automatic control system. This model can also be on the aircraft landing data collected at Chicago's O'Hare airport, used to solve runway exit location problems by providing the expected the model divided the landing process into three phases. Phase 1 distribution of aircraft landing distances and predict aircraft runway accounts for the aircraft motion from threshold to the touchdown occupancy times. An interactive computer program has been developed point with the vehicle flying at a constant speed profile. In phase 2 on an IBM RISC 6000 workstation to perform these tasks. the aircraft uses a deceleration rate consistent with the use of reverse thrust until it reaches a coast speed. Phase 3 has two cases in which With the increase in air traffic demands, airport ground network the aircraft either uses the deceleration rate in Phase 2 to reach an operation analysis becomes more important to fully realize the exit with the required turnoff speed (i.e., there is an exit located at capacity of airports. The use of new Air Traffic Control System that place) or the aircraft coasts for ~Ttime and then uses the decel­ (A TC) technologies in the near future could reduce the aircraft in eration rate in Phase 2 to reach the exit with the required turnoff trail separations in the airport terminal area thus making aircraft run­ speed. This model does not consider any influence of airport layout way occupancy times become an important factor in determining air­ and environmental factors such as the gate locations, runway port capacity. The expected intensity of runway operations in the grades, the weather conditions, and so on, which may cause signif­ future will also influence the safety of these operations thus requir­ icant deviations in aircraft landing operations at different airports. ing more precise methods of determining aircraft state variables in Several empirical studies on aircraft landing behaviors were con­ real-time on a ground network. Landing aircraft processing is one of ducted in the late 1970s (4,5). Through analysis of observations col­ the key factors in airport ground network operation analysis. A bet­ lected at different airports, Koenig (4) found that a key factor influ­ ter understanding of the aircraft landing process could help to encing the aircraft selection of an exit is the terminal gate location. improve airport ground operation management and the safety of air­ Other factors such as the traffic density, passenger comfort, and so craft operations. Furthermore, it could provide knowledge for ground on, also have influence on the landing performance. He found that network designs including the optimal runway exit location problem. pilots in many instances have the motivation to exit early in order The Aircraft Landing Simulation and Prediction Model to reach their assigned gate in shorter times'. He pointed that this (ALSPM) described here has been calibrated using real aircraft motivation factor could be used to reduce runway occupancy times. landing data observed at five major airports in the United States. In 1990, Ruhl ( 6) presented an aircraft landing model which uses With real-time in~ut data, this model could provide landing infor­ aircraft individual parameters to predict the aircraft runway occu­ mation instantly to both pilots and air traffic controllers in a ground pancy time. In this model, aircraft runway landing operations are traffic automatic control system. The information could include divided into five segments. In Phase 1 the aircraft crosses the thresh­ acceptable runway exits, the probability of each aircraft taking these old and travels at a constant speed until a flare maneuver is initiated. exits, related runway occupancy times, and advisories on the short­ Phase 2 encompasses the flare maneuver and ends when the main gear est taxiway path to an assigned gate. The model described here can touches down. Phase 3 starts from the point where the main gear touches down until the nose gear impacts the ground. Phase 4 starts at the nose gear touchdown point with the aircraft speed bleeding off Department of Civil Engineering, Virginia Polytechnic Institute and State at an average braking deceleration rate until reaching a suitable exit. University, Blacksburg, Va. 24060 It is assumed here that if the deceleration rate under normal condi- Gu eta/. 53 tions allows the aircraft to accept an exit (means the aircraft could landing aircraft. Motivation factors, runway, and weather conditions decelerate to the required speed before it reaches the exit) the pilot factors are considered in the simulation to represent different airport will adjust (decrease) the deceleration rate to meet the required exit environments. The most important improvement in this model is the speed at the time the aircraft reaches the exit. Phase 5 starts at the consideration of the gate location as a motivation factor. A multi­ point the aircraft begins to tum off on the runway until it clears the objective optimization method is implemented here to link this fac­ runway. One shortcoming in this model is the obvious simplification tor to the aircraft landing performance and exit choice. The model of the aircraft deceleration phase (Phase 4). According to our obser­ describes the aircraft landing roll performance based on the consid­ vations, pilots use different strategies in this phase based on the exit eration of a complete ground network. This provides more realistic location. For example, an aircraft may decelerate to a certain speed results which could be used in automatic ground control system and coast for some time and decelerate again to reach its exit. Sim­ development and runway exit location optimization procedures. plifications in this phase usually result in higher runway occupancy times than those observed in the field. Another problem is that there is no inclusion of motivational factors in this model. Ruhl mentioned Aircraft Landing Process Description the influence of the terminal location to the aircraft landing operation in his paper. However, the model did not consider this factor (6). The aircraft landing process is broken down into five phases: flare Another aircraft landing simulation model was developed to esti­ phase, first free roll phase, braking phase, second free roll phase, mate aircraft runway occupancy time for runway exit location and and turnoff phase as illustrated in Figure 1. runway occupancy time minimization at Virginia Polytechnic Insti­ The flare phase starts from runway threshold until the aircraft tute (7). This model also divided the landing process into five touches down. The landing distance, Sain and travel time, tain are esti­ phases including a flare phase, two free roll (or transition) phases, a mated by Equations 1 and 2 under the assumption that the aircraft braking phase, and a turnoff phase as shown in Figure 1. Several uses a steady descent flight path angle g (3.0° typical) with a constant random variables, such as the approach speed, aircraft landing nominal acceleration during the flare maneuver at 1.2 g's (9, JO). weight factor, and the deceleration rate during the braking phase, are generated using a Monte Carlo sampling technique. Factors that ~ Vi have influence on the aircraft landing operation are included in this Sair = + ( · + M(rl) (1) 'Y 2 g n - 1) model, such as weather conditions and the local effect of runway 0 grades. Runway length as a pilot motivation factor is also consid­ 2Sair ered and a more realistic braking phase related to the aircraft exit fair= (2) choice is used. However, the gate location influence was not con­ sidered to simplify the complexity associated with a runway exit where optimization model and its portability on a personal computer (7,8). This study addresses some of the limitations of previous models h h = threshold crossing height, and describes a technique to predict landing roll performance in 1 Vn = flare speed, realistic airport operational conditions considering gate location as nn = flare load factor, a causal factor in the exit choice model. 11S(rl) = adjustment distance of Sair according to different runway length (rl), Var = aircraft approach speed, and MODEL DEVELOPMENT vtd = touchdown speed.

Based on the five-phase aircraft landing process shown in Figure I, The first free roll phase starts at the point where the main gear the model uses Monte Carlo simulation to perform 250 trails for each touches down and ends when thrust reverses and braking are

Exit Decision Turnoff Entrance Touchdown Point Point Point Aircraft Speed

Approach Speed

Decision Speed Exit Speed

FIGURE 1 Aircraft landing phases. 54 TRANSPORTATION RESEARCH RECORD 1506 applied. It is assumed that aircraft travels at a constant speed for An exit choice model is used in the braking phase to determine the about 1-2 sec. most likely exit to be used. This model will be described later. The second free roll phase is scheduled after the braking phase (3) just before the aircraft starts turning off from the runway. This phase is associated with the pilot identification and decision procedure to where sfrl is the first free roll distance, and f1 is the travel time. take a specific exit. The aircraft will travel at a constant speed (i.e., The braking phase starts from the ending point of the first free roll the exit speed) for about 1-3 sec. phase until the aircraft decelerates to an acceptable exit design speed (Vex). The aircraft uses a nominal deceleration rate to decelerate to a (7) speed called decision speed (Vies). The model checks for a possible coasting distance (Scoas1), under the assumption that the aircraft uses where the nominal deceleration rate to reach the selected exit after coast­ ing. If this distance is within certain range (Idec), the aircraft uses the Srr2 = second free roll distance, adjusted deceleration rate to reach the exit's design speed without t2 = travel time, and coasting. If the distance exceeds Idec. the aircraft coasts for some time Vex = exit speed. under the decision speed and then uses the nominal deceleration rate to decelerate to the exit. The nominal deceleration rate (dee), is cal­ The turnoff phase is used to describe aircraft exit turnoff behavior culated considering the ma.nufacturer' s published landing distance and estimate the turnoff time. This phase starts from the point where and subtracting an air distance (7). It is also adjusted by runway aircraft begins the turnoff maneuver and ends at the point where the local gradient, surface conditions (wet or dry), aircraft landing aircraft clears the runway. The turnoff time (t10r) is estimated weight information and the aircraft assigned gate location. The deci­ through numerical integration using a 4th order Rung-Kutta algo­ sion speed used in the model has been obtained through empirical rithm (12) as shown in Equation 8. data collected at various airports (11). Equations 4 and 5 are used to estimate the braking phase distance (Sbr) and time (tbr). (8)

(4) where

biail = aircraft tail plane span, if Scoast < Idec hwing = aircraft wing span, Rw = runway width, and ET = selected exit type. (5) v~ - Vix Sbr - Vid - Vdec 2 X dee As described above, the aircraft runway occupancy time ROT can + if Scoast 2 f ctec dee Vdec be estimated by adding all individual times in all phases. where lex is the distance from a selected exit to the runway thresh­ ROT = fair + f1 + fbr + f2 + f1of (9) old, and Scoast is the possible coasting distance which can be calcu­ lated by using Equation 6. Figure 2 shows two Boeing 727-200 landing simulation trajecto­ V~ct-v;x) ries to illustrate differences in landing roll behavior at two hypo­ Scoast = lex - Sair Srr1 Srr2 2 X dee (6) ( + + + thetical runway exit locations.

Exit ROT (s.) Location (m.) • 1700 53.5 I '){)()() 61 1 70- t .-..... • • • 111 Decision Cl) 60-. 1. Point s f ••••• • •. • / Adjusting Braking "'O 50- 0 . Flare Deceleration Phase 0 1 •""' 0. 40- Phase •., ...... _ j ~ ~ Nominal Braking •·• ._~ ~ .....clS 30. u . Phase •._.. ~ 20- ..... , . 10 0 25o 500 15o 1060 i 25o 1s0o 1150 2000 Distance from Threshold (m)

FIGURE 2 Sample velocity profiles (Boeing 727-200). Gu et al. 55

Aircraft Stochastic Landing Behavior Figures 3 and 4 show Boeing 727-200 landing distributions (to a speed of 15 m/sec) simulated by the model for two values of the Monte Carlo sampling technique is used in runway landing simula­ landing motivation factor. The terminal location for this example is tion to represent the stochastic behavior of landing aircraft. Created assumed to be near the active runway threshold (i.e., pilots could be random variables include aircraft landing weight factor (wr) (7), heavily motivated to shorten their landing rolls to reach the termi­ runway crossing height (hr11), flight path angle (-y), flare speed (V11 ), nal location). deceleration rate in braking phase (dee), and exit speed (V.x). All of these random variables are assumed to have normal distributions as shown in Equation 10. Strategy 2

X ~ N(µ, rr) (10) The shortest aircraft taxiing time to the terminal location is used as a factor to influence aircraft exit choice. It is assumed that the land­ where ing aircraft will choose the acceptable exit which can minimize its runway occupancy time plus its weighted shortest taxiing time (see X = random variable, Exit Choice Model in detail). µ = mean value, and rr = standard deviation. The Exit Choice Model The upper and lower boundary values are set for each distribution according to the landing observation analysis carried out by the Vir­ A multiobjective integer optimization model is developed to find ginia Polytechnic Institute Transportation System Laboratory at five the exit for landing aircraft. Minimizing the aircraft runway occu­ east-coast airports (11). pancy time (ROT) and minimizing taxiing time (Tr) are the two objectives. A taxiing time weight factor is used to combine these two objectives. The following two assumptions are made in the Terminal Location Influence on Landing Process model: 1) The landing aircraft will choose the acceptable exit which can minimize its ROT plus its weighted TT. 2) The aircraft ROT is The terminal location is known to have influence on the aircraft run­ at least equally important with TT. This fact is used to achieve a bal­ way landing behavior (4,6,11). In this model, two strategies are used ance between individual and collective (system wide service times). to reflect this influence. The model can be described mathematically as follows:

Strategy 1 11 Minimize L (ROT;k + w_h + TT;k)x; If under the nominal deceleration rate an aircraft passes over the i=I perpendicular plane of the terminal location, a more aggressive 11 deceleration rate is used by the model to reflect the pilot's motiva­ Subject to L X; = 1 (11) i=I tion for attempting an earlier exit. lex(i) ?:. Sair + Srr1 + Srr2 + Sbr deCnor if Lr

decnor = nominal decoration rate, i = runway exit index number, -y = landing motivation factor, n = total number of exits, Lr = aircraft nominal landing distance, and lexCi) = location of the ith exit, f(GL) = function of terminal location GL. k = terminal index,

14 12 QI) "'Q :a 10 § ...J 8 ...... 0 ..... (\) 6 ..c E 4 z;::) 2 0 900 1000 1100 1200 1300 1400 1500 1600 Landing Distance (m)

FIGURE 3 Boeing 727-200 landing distribution (A = 0). 56 TRANSPORTATION RESEARCH RECORD 1506

14

rn 12 OJ) :ac: 10 § -l,_ 8 0 ..... 6 11) 8 4 z::3 2 0 900 1000 1100 1200 1300 1400 1500 1600 Landing Distance (m)

FIGURE 4 Boeing 727-200 landing distribution (A.= 0.1).

W.h = taxing time weight factor for airline k, and 3. Aircraft landing simulation: two hundred and fifty landings are X; = binary variable which indicates the aircraft will either generated using a Monte Carlo sampling technique. The model gen­ take exit i (1) or not (0). erates random trajectories according to variations in the runway crossing height, flight path angle, landing weight factor, aircraft Since the number of runway exits is limited, we use a numerical deceleration rate, and exit speed. Each landing follows the five land­ method to solve this integer program problem. The determination ing phases described before. of wfk will be discussed in following section. 4. Simulation results: by sorting the simulation results of each landing and recalculation, the model will provide the exit choice Taxiing Time and Taxiway Path Prediction probabilities to each exit, runway occupancy times, and the short­ est taxiway path. The aircraft shortest taxing time and taxiway path can be estimated by solving the following shortest path optimization model. Suppose we have a taxiway network G with m nodes, narcs, and APPLICATION a cost cij associated with each arc (i,j) in G. The shortest path prob­ lem is to find the least costly path (i.e., shortest path) from node i to Operations at Washington National Airport have been studied to node). ,, test the validity of this model. This airport has also been used to see The mathematical description is to find the shortest path from the possible effect of terminal location on landing roll performance. node l to node k: Based on this application, sensitivity analysis has been done by

111 111 changing the terminal location and the value of taxing time weight Minimize I I cijxij factor. i=I j=I As mentioned before, the model could also be used in a ground Ill 111 {l if i = l traffic automatic control system. With real-time information such as Subject to Ixij - Ixk; = 0 if i_ ~ l or k (12) the aircraft approach speed and the touchdown distance which could 1= I k= I - 1 If l = k be provided by airport radar system or on board equipment, the model could serve as an advisory system to pilots and air traffic con­ xij = 0 or 1 trol personnel to automate the aircraft runway exit choice and find i,j =I, ... , m the shortest path to an assigned gate thus saving fuel and minimiz­ ing runway occupancy times. Model predictions with known where xij is a binary variable which indicates arc (i, )) is either in the approach speed (i.e., extracted from radar or differential GPS trans­ path ( 1) or not (0), and cij is the travel time that the aircraft spends mitters) could also reduce the standard deviation of runway occu­ on link (i, )). To simplify this model at this stage we assume that the pancy times thus contributing to a better utilization of runway infra­ aircraft taxiing speed is constant on each taxiway link. structure.

Procedures in Landing Simulation and Prediction 1. Case Study The landing simulation and prediction include following steps: 1. Airport environmental data input: these data include informa­ We used runway 36 at the Washington National Airport as an exam­ tion on runway and taxiway network, the prevailing airport weather ple to test the proposed model. The length of this runway is 2040 condition, and terminal locations. meters and the runway exit information is shown in Table 1. 2. Taxiing time weight factor determination: to calibrate wfb air­ Figure 5 shows the runways and related taxiway network at this craft landing roll data should be collected and analyzed. Running airport. Landing events for USAir Boeing 737-300 aircraft are pre­ the model for different values of taxiing time weight factor wfk, and dicted and compared with field observations collected at the airport. comparing the results with field data, we can find the best suitable The selection of one airline and one aircraft model is made to nar­ taxiing time weight factor. row the scope of the possible correlation of parameters in the model. Gu etal. 57

TABLE 1 Runway 36 Exit Information Exit Exit Location From Exit Speed Name Type Threshold (m) (m/s) G Pseudo 90 Deg. 950 10 H Pseudo 45 Deg. 1025 15 I 45 Deg. 1325 15 RWY Pseudo 30 Deg. 1470 18 J 90 Deg. 2040 10

Washington National Airport (Drawing not to scale)

ExitJ

9

US Air. Terminal Location it 10 Terminal Location for Sensitivity Analysis FIGURE 5 The ground network of Washington National Airport.

The USAir terminal is also particularly suitable in this analysis (0.4), we have the different predictive results. Figure 8 shows the because its location is near the end of runway 36. exit choice predictions for three different terminal locations. Preliminary analysis of 36 observed Boeing 737-300 landings It is obvious that the taxiing time weight factor has no influence found that the most suitable taxiing time weight factor for this air­ on the landing predictions when the terminal location is at Node 10. craft/runway combination to be 0.4. Figure 6 shows the exit choice This is because minimizing the aircraft runway occupancy time will predictions and the observed exit choice distribution. Figure 7 minimize the aircraft taxiing time automatically in the exit choice shows the runway occupancy time prediction and the observed val­ model. In this case the difference is attributed to changes in the ues. Note that in both cases there is good agreement between pre­ landing motivation factor. Landing aircraft use more aggressive dicted and observed values. deceleration behavior in the braking phase. The model prediction is sensitive to the airline terminal location. The model predictions are also sensitive to the value of taxiing Hypothesizing different terminal locations (indicated as nodes 9 time weight factor. Figure 9 shows the exit choice predictions under and 10 in Figure 5) and using the same taxiing time weight factor different taxiing time weight factor values (with the terminal loca-

80 ~ 60 ~ [ill Real Data ~ ..0 0 1-o Iii) Prediction p... 40 QJ u "8 ..du 20 ·;;:;: UJ 0 H RWY J Exit Name

FIGURE 6 Exit choice observation and prediction. 58 TRANSPORTATION RESEARCH RECORD 1506

...... 50 ~ d) s 40 ~ >. (.) [fill Real Data ~ 30 0.. :;) Prediction (.) D (.) 20 0 >. CIS ~ 10 Q :;) ~ 0 H RWY Exit Name

FIGURE 7 Runway occupancy times observations and predictions.

100

~ Assigned Gate .~ 75 Locations:

.0~ [ill Location at node 8 0 50 ~""' [2] Location at node 9 Q) ·o(.) ~ Location at node 10 u.d 25 ·~ tI.l 0 H RWY J Exit Name

FIGURE 8 Influence of terminal location on exit choice probability. tion at Node 8). We can see that the landing aircraft will have a ten­ state variable information extracted from aircraft surveillance radar dency to take the exits closer to the terminal when the taxiing time data or information download from on-board navigation equipment weight factor is increased. This is representative of motivated pilots and differential position sensors on the ground. Figure 10 shows the who are willing to trade off some runway occupancy time (ROT) for expected landing distribution for Boeing 737-300 aircraft with no taxiing time. This fact, when taken to an extreme, could affect the real-time update data. Figure 11 hows the landing distribution with runway acceptance rate due to longer ROT values. known approach speeds (55 m/sec) and touchdown point location (450 m). We can see from these two figures that with more infor­ mation the aircraft landing behavior tends to be more predictable 2. Prediction With Real-Time Data and the resultant dispersion of runway occupancy times could result in slightly higher runway acceptance rates. This could prove to be With real-time data updates this model could provide more accurate useful if in-trail separations under Instrument Meteorological Con­ prediction . Automated update could come in the form of aircraft ditions (IMC) are reduced further from current values as the mag-

80 ~ 0 Weight Factor ~ 60 ~ [ill wf=O .0 0 lS}l wf=0.4 ~""' 40 ~ wf=0.8 Q) .$:? 0 u.d 20 ..... ·~ tI.l 0 H RWY Exit Name

FIGURE 9 Influence of taxiing weight factor. Gu et al. 59

10 Aircraft Model: Boeing 737-300

~::s z 2

0 -+--,.~-r-----..L-,.._-+-___,'--+----'r--'T-.....,__._,...... --.--....._,_._.,...... ----r....._~ 900 950 1000 1050 1100 1150 1200 1250 1300 Landing Distance (m)

FIGURE 10 Landing distributions with no real-time update data.

nitude of runway occupancy times will be closer to in-trail separa­ tional safety. The model results could also be used in the runway tion. Also, under current Visual Meteorological Conditions (VMC) exit location problem to increase runway capacity. with heavy banks of flights, a reduction in the mean and standard Further studies are needed to improve the model prediction capa­ deviation of ROT could enhance the safety of operations by increas­ bility, including the influence of traffic density on aircraft landing ing the gaps between successive arrivals. behavior, the determination of landing motivation factor for a mul­ titude of airport/aircraft combinations, and their associated taxiing time weight factors. Also, the development of a time dependent traffic assignment algorithm would help in the predictions for an CONCLUSIONS AND FURTHER RESEARCH advanced ground control A TC system.

The preliminary results presented in this study indicate that the model described provides a more realistic way to analyze landing aircraft behaviors in complete airport ground networks. The ACKNOWLEDGMENTS use of individual aircraft parameters and a calibrated model using observed landing data made the analysis more accurate. The authors gratefully acknowledge the support of this research The most important improvement in this model is the considera­ by the Federal Aviation Administration under the NASA Con­ tion of terminal location influence on landing behavior. This con­ tract NASl-18471 Task No. 15 and the continuing support by sideration makes it possible to explain the aircraft landing roll phe­ the Operations Research Office of the Federal Aviation Admin­ nomena consistent with a particular ground network with more istration under the Contract 93-G-067. The advice and guid­ accuracy. ance of Hisao Tomita, Satish Aggarwal, Jim White, and Steve Providing landing predictions to the pilot and air traffic controller Bradford at the FAA and David Middleton at the NRCL, who in the ground traffic control system, the model could help to solve acted as technical monitors on these research projects, are greatly ground network traffic congestion problems and enhance opera- appreciated.

10 +-...... ~.....__._~..__...... ___.~....__._~..__...... ____..___.___.,~...... ___._--+ Aircraft Model: Boeing 737-300 9 Approach Speed: 55 mis TD. Distance: 450 m 6 Final Speed: 15 mis 'O 5 ..... ~ 4 § 3 z 2 1 0 -+-----r~...---,-~..-----.---,.-'-+-...... _._~...... ~~...... L...... J...... --+ 900 950 1000 1050 1100 1150 1200 1250 1300 Landing Distance (m)

FIGURE 11 Landing distributions with known approach speed and touchdown distance. 60 TRANSPORTATION RESEARCH RECORD 1506

REFERENCES 7. Trani, A. A., A.G. Hobeika, B. J. Kim, and V. Nunna. Runway Exit Designs for Capacity Improvement Demonstrations: Computer P~o­ gram Development. U.S. Department of Transportation, Federal Avia­ 1. Horonjeff, R., D. M. Finch, D.S. Belmont, and G. Ahlborn. E~it Tax~­ tion Administration, Report RD-92/I-32, Washington, D.C, 1992. way Location and Design. Institute of Transportation and Traffic Engi­ 8. Hobeika, A. G., A. A. Trani, H. D. Sherali, and B. J. Kim. Microco~­ neering, University of California, Berkeley, Calif., 1958. . puter Model for Design and Location of Runway Exits. Journal of 2. Daellenbach, H. G. Dynamic Programming Model for Optimal Loca­ Transportation Engineering, Vol. 119, No. 3, 1993, pp. 385-401. tion of Runway Exits. In Transportation Research, Vol. 8. Pergamon 9. Roskam, J., and E. Lan. Airplane Aerodynamics. Roskam Publishing Press, New York, 1974, pp. 225-232. Co., Lawrence, Kans., 1981. 3. Joline, E. S. Optimization of Runway Exit Configurations. Journal of 10. Toreenbeek, E. Synthesis of Subsonic Aircraft Design. Van Nyhus Pub­ Transportation Engineering, Vol. 100, 1974, pp. 85-102. lishing Co., 1981. 4. Koenig, S. E. Analysis of the Runway Occupancy Times at Major ~ir­ 11. Transportation System Laboratory. Landing Data Collected from Five ports. Report FAA-EM-78-9. FAA, U.S. Department of Transportation, Airports. Civil Engineering Department, Virginia Polytechnic Institute, 1978. Blacksburg, Va., 1993. 5. Jackson, T. J., and E. P. Moy. An Analysis of Runway Occupancy Time 12. Trani, A. A., A.G. Hobeika, H. Tomita, and D. Middleton. Character­ for Large Aircraft at San Francisco lnternation~l Airpo:t. R~port UC~­ istics of High Speed Turnoffs. Presented at Proceedings of the 22nd ITS-GR-79-3, Institute of Transportation Studies, Umvers1ty of Cah­ International Air Transportation Conference, Denver, Col., 1992. fornia, Berkeley, 1979. 6. Ruhl, T. A. Empirical Analysis of Runway Occupancy with Applications to Exit Taxiway Location and Automated Exit Guidance. Report 1257, Publication of this report sponsored by Committee on Airfield and Airspace TRB, National Research Council, Washington, D.C., 1990, pp. 44-57. Capacity and Delay. TRANSPORTATION RESEARCH RECORD 1506 61

Economic Characteristics of Multiple Vehicle Delivery Tours Satisfying Time Constraints

MAX K. KIESLING AND MARK M. HANSEN

Since deregulation of the aviation industry, a substantial body of litera­ Northwest Airlines, United Airlines, and USAir) have joined forces ture has emerged analyzing the economic structure of passenger carrier with Roadway Package System (RPS), a national ground carrier, to operations. By comparison, a paucity of literature exists that addresses provide door-to-door delivery of freight and compete with inte­ the economics of air freight transportation. This study contributes to fill­ grated freight carriers. Such an alliance has far-reaching implica­ ing that void by assessing the economic structure of ground-side freight distribution for air express carriers. To do so, we develop an "engineer­ tions. If passenger carriers can effectively compete with dedicated ing" cost model of the ground-side distribution process. This circum­ freight carriers, there will be a continued need for joint-use airports vents the problem that appropriate historical performance data is not (i.e., both freight and passenger carriers using the same airports). If available with which to develop an "economic" cost model and affords they cannot compete, there will be an increased specialization of greater flexibility and accuracy than the more frequently applied econo­ services (specializing either in freight or passenger transportation) metric based cost models. The cost model is developed by first employ­ and, consequently, an increased need for specialized airports. ing a mathematical heuristic to design and locate freight delivery subre­ gions employed by freight carriers operating under time constraints. The To assess how competitive dedicated freight carriers and combi­ results of the design heuristic are then used to create a model that incor­ nation carriers can be (from an economic perspective), we must be porates costs of overcoming distance, stopping costs, marginal freight able to quantify the operational cost of freight delivery. One distribution costs, and fixed vehicle costs. It is then used to demonstrate accepted way to do this is to develop an "economic" cost model that ground-side freight distribution operations exhibit significant using industrywide or carrier specific data to calibrate econometric economies of scale and profound economies of density. Furthermore, it cost and/or production functions [Kiesling and Hansen (4)). Unfor­ is indicated that increasing the deli very time constraint decreases distri­ tunately, data are very limited, particularly for dedicated freight car­ bution costs. However, this decrease in costs must be tempered with the trade-off that increasing the delivery time constraint could decrease the riers and for specific delivery operations such as ground-side distri­ market available to the carrier. bution, to support such an analysis. As a result, we employ another possible approach, which is to develop an "engineering" cost model of the more specific operations of air freight carriers. The results The air cargo industry was deregulated in November 1977, 1 year presented in this paper are the first step in developing such a model. before deregulation of the passenger airline industry. At that time, (Whereas the final goal is to develop engineering cost models of air cargo was primarily transported in the bellies of passenger air­ system-wide operations, this study addresses only ground-side craft with the notable exception of the cargo transported by Flying transportation costs.) Tigers, a successful international air freight forwarder. Door-to­ Ground-side pickup and delivery operations are a crucial battle­ door delivery was uncommon, and overnight delivery was the ground in the competition between specialized and combination exception, not the norm. freight carriers. One reason is that pickup and delivery operations The industry changed dramatically after deregulation, as Federal are the interface wherein customers judge the level of service Express Corporation, a small package express carrier, emerged and received. As delivery deadlines attest, one critical factor in defining rapidly grew to dominate the air freight industry. Federal Express' the level of service is time. Air express customers pay premium rapid growth eventually led to their purchase of Flying Tigers [see rates for the timely transport of goods, both in the sense that pickup Sigafoos (1) and Trimble (2)). Attracted by Federal Express' rapid and delivery deadlines are reliably met, and in the more general rise to dominance and success, several other specialized air freight sense that freight can be delivered as early as possible in the busi­ carriers emerged including UPS, Airborne Express, and OHL. By ness day and be picked up as late as possible in the business day. the mid-1980s the air cargo industry was dominated by these ser­ The determination of pickup and delivery deadlines is one decision vice oriented carriers, forming the organizational structure that variable that effects the level of service provided. As will be indi­ exists today in the aviation industry: specialized carriers that focus cated in this report, however, it is also a decision variable that sig­ on either cargo or passenger transportation. nificantly effects the costs incurred by the freight carrier-shorter One factor contributing to the rapid rise of dedicated air freight time constraints raise the operational costs to the carrier. This trade­ carriers was the apathy of passenger carriers toward air cargo fol­ off will prove to be a crucial element in the competition between lowing deregulation. However, there is reason to believe that the freight carriers. passenger carriers' apathy has ended. Shaw (3) reports that five The design and operation of multiple-vehicle delivery systems major U.S. passenger carriers (American Airlines, Delta Air Lines, (such as those described above) have been analyzed by numerous authors. Daganzo (5) explores the impact that zone shape has on The University of California at Berkeley, Institute of Transportation Stud­ tour building strategies and ultimately on tour lengths. Daganzo (6) ies, !07 A McLaughlin Hall, Berkeley, Ca. 94 720-1720. presents a strategy for designing distribution problems in which N 62 TRANSPORTATION RESEARCH RECORD 1506 points must be visited by a fleet of vehicles operating under the con­ pickup and delivery points in the subregion). For a given number of straint of a maximum of C stops per vehicle. Newell and Daganzo pickup and delivery points in a city, increasing N by one increases (7,8) expand this work further by considering larger delivery areas the total line-haul distance traveled by an amount on the order of the wherein line-haul distances are significantly greater than local travel average distance from the terminal to all pickup and delivery points distances. Newell (9) modifies the analysis to consider the move­ in the city. However, it decreases the total local distance traveled by ment of valuable goods. In all of the above studies, vehicles are con­ approximately the average travel distance between pickup and deliv­ strained by capacity. Relatively little has been done on the design ery points. Because line-haul trips are almost always much longer of multiple vehicle delivery systems constrained by time. One such than the average distance between pickup and delivery points, it fol­ study, by Langevin and Soumis (10), does consider this problem, lows that the total travel distance increases with N. Thus, to minimize but only for ring-radial networks and a centrally located depot. Han the cost of overcoming distance, we would want to minimize N. (11) also focuses primarily on ring-radial networks in developing The relationship between N and the labor cost of delivery follows routing strategies for multiple vehicle delivery problems. a similar vein. Let T1 be the total labor (hours) required to service This study explores the design of multiple-vehicle delivery sys­ all pickup and delivery points in the city, which is comprised of the tems constrained by time (not vehicle capacity or dispatch fre­ total time required to travel (both line-haul and local travel) and the quency), and applies the design process to several different types of total time required to handle and process freight at each pickup and cities. Section l presents the basic distribution process that is delivery point. The latter is constant regardless of the size of N. employed by air freight carriers, and defines the basic' design prob­ Since the total travel time is directly proportional to the total dis­ lem. Section 2 applies the design process to linear cities and cities tance traveled, it is obvious that the total travel time also increases with Li metrics. Section 3 discusses the impact of "fast roads" on the with N. Thus, the labor costs are comprised of fixed and variable design of delivery subregions, and extends the design process to (with N) components, which are minimized by minimizing N. allow for fast roads. Section 4 uses the results from Sections 1-3 to Finally, it is clear that if one delivery vehicle is assigned to each estimate the average unit cost of transporting freight on the ground­ subregion, the vehicle cost is also minimized by minimizing N. side distribution system, and demonstrates the crucial role that These transformations allow the cost function to be rewritten: pickup and delivery time constraints play in ground-side distribution. (2)

MINIMIZING DELIVERY COSTS where f! (N) 2::: 0. Therefore, to minimize costs, carriers should min­ imize the number of delivery subregions required, subject to the Simply stated, the air freight carrier's goal in designing its ground­ constraint that all points are visited in time T. side distribution system is to visit all pickup and delivery points in the city at minimal cost. Two constraints determine how many vehi­ cles are required to accomplish this task. First, vehicle weight and DESIGN OF MUL TIVEHICLE DELIVERY ZONES volume constraints are likely exceeded before all points in a city can be visited, even in small cities. It follows that the next best solution Designing delivery subregions is a detailed, and case specific, activ­ is to fully use delivery vehicles by visiting as many points as possi­ ity. Results differ with changes in the terminal location or the under­ ble before weighing-out (meeting the vehicle's weight limit) or cub­ lying transportation metric. The design process remains the same, ing-out (meeting the vehicle's volume limit). However, the time­ however, as formalized below. sensitive nature of delivery deadlines precludes the vehicles from Let Tbe the amount of time allotted to visit all points in the deliv­ even visiting enough points to reach vehicle capacity, meaning that ery area (city). Only one delivery vehicle visits each subregion in time is the second, and as it turns out the binding, design constraint. time T. For the delivery process, T includes the time required to The solution is to divide the delivery area into subregions, the size travel to the delivery subregion and visit all points in the subregion. of which are determined by the maximum number of points that a For the pickup process, it includes the time to visit all points in the single delivery vehicle can visit in the allotted time. This is equiva­ subregion and return to the terminal. (For cost estimating purposes, lent to minimizing the number of delivery subregions in the deliv­ both line-haul trips must be included.) ery area. . We can analytically express the design constraint by defining To illustrate analytically, consider the following simplified three time quantities: the line-haultime, Tl, which is time required ground-side delivery cost function facing a freight carrier: to travel from/to the terminal to/from the delivery subregion; the handling time, T2, which is the time required to transport freight (I) to/from the customer from/to the vehicle; and the local travel time, T3, which is the time required to travel the local streets between where pickup and delivery points. The sum of these activities must be less than or equal to T for all delivery subregions: Cd = cost per mile traveled,

C1 ~ cost per hour of labor, and Tl+T2+T3~T (3) C,. = fixed cost per vehicle. This basic constraint holds true for all transportation metrics and First, consider the cost per distance term of Equation 1. Let N be city shapes analyzed in the remainder of this section, wherein several the number of delivery subregions required to visit all pickup and different scenarios are analyzed. For simplicity of demonstration, a delivery points, n, in a city. Each delivery tour consists of a line-haul linear city is analyzed first. The design process is then expanded and portion (the distance from the terminal to the nearest point in the applied to cities with Li transportation metrics, a scenario that is tour), and a local travel portion (the distance required to visit all much more realistic than linear or ring-radial cities. Finally, the Kiesling and Hansen 63

impact of fast roads on the design of delivery subregions is consid­ It should be noted that d 1, and all remaining calculations of d;, can ered, providing the most realistic design framework possible. be solved in this manner only by assuming that time is the binding

constraint. By substituting [De - (d2 + d 1)] for D; and d2 for d; in the time constraint, di is easily determined. Solving recursively, an Linear City expression emerges that allows us to design zone i (for i = 2, ... , n - 1) where n is the total numbers of zones required to cover the To formalize and demonstrate the design process, a linear city of entire city: length De is considered first. A terminal is located at one end of the i city, and all points to be visited are randomly distributed across its Tv - De+ Ld;_, length. Delivery subregions are nonoverlapping zones of length d;, di= ~~~~~j_=_2~~ (6) located a distance D; from the terininal, as shown in Figure 1. Sub­ 8-rv regions are located adjacent to one another, so that 'i.di = De. The line-haul time, Tl, is the time required to travel between the The zone adjacent to the terminal, zone n, is simply the remaining terminal and the nearest edge of the subregion. By assuming an length of the city, average velocity, v, the line-haul travel time to subregion i is sim­ 11 ply Tl = D; .,. 1 d,,* - De - L~ dj-1* The handling time, T2, for a vehicle of the ith subregion is the time j=2 required to perform the delivery or pickup tasks at all points in the subregion. Such a task includes parking the vehicle, walking to the It will be less than the length determined by the above design appropriate location, processing the required paper work, handling equation. the package, and returning to the vehicle. To obtain the total handling Thus, given a linear city of length De and customer density B, we time in subregion i, we assume that the handling time per stop, T, is can optimally design all subregions. We simply begin with the constant on the average, which we then multiply by the total number above expression for df, which gives the optimal size and location of points in the subregion. If 8 is the customer density (number of of the outermost subregion. Then, knowing the length df, we can points per unit length), then the expected number of points in subre- determine the length of all other zones (i = 2, ... , n - I) recur­ . gion i is 8 d;. Thus, the total handling time of zone i is T2 = 8d;T. sively with the above expression ford( . The third element of the time constraint is the local travel time, which is the time required to travel between all points in a specific

zone. When the number of points in a subregion is sufficiently large, L 1 Metrics the distance traveled is closely approximated as the length of the subregion. If there are few points in the subregion, however, it may A linear city is clearly an unrealistic representation of any city that be deemed necessary to reduce the travel distance by one half the would be included in air freight networks. However, the design expected distance between points, 1/(23). Assuming there is a suf­ process that applies to the hypothetical linear cities also applies to ficiently large number of points in the subregion, the local travel two dimensional cities. Since U.S. cities rely primarily on rectan­

time in subregion i is T3 = d;;,,. gular (L 1) transportation metrics, we need to adapt the design Having defined all three tasks, the time constraint facing vehicles process to apply to such metrics. In the following pages, we apply

in subregion i can be rewritten: the design approach to cities with L1 metrics when the terminal is located in the city center, on the edge of the city, and in one corner D· d· of the city. --i- + 8d;T + -¢ ~ T (4)

In designing the subregions, the underlying goal is to minimize Terminal in City Center the number of vehicles required, which is equivalent to maximizing the number of points per subregion. The design process is begun by First, consider a delivery area with a centrally located terminal, as considering the outermost delivery zone, subregion 1. Its optimal shown in Figure 2(left). For analysis purposes, the delivery area length, dt, is determined by replacing the line-haul distance, D;, in shape is approximated as a square oriented at 45° to a fine orthogo­

the first term of the time constraint with the line-haul distance to the nal transportation grid (L 1 metric), a shape dictated by the equi­ first zone, De - di, and solving for d 1: travel time contours (the locus of all points that can be reached in a given amount of time). The size of the delivery area is defined by Dn which is the travel distance to the outermost corner or edge of d* = Tv - De (5) I OTV the city. All points to be visited are distributed randomly through-

FIGURE 1 Subregion design for a linear city with one terminal on edge of city. 64 TRANSPORTATION RESEARCH RECORD 1506

FIGURE 2 Subregion design for city with an L1 metric and centrally located terminal.

out the area with a constant density, 3(x, y) = 3. Vehicles travel at The local travel distance is approximated as the product of the speed v throughout the city. number of points in the subregion, defined above, and the expected The first step in designing delivery subregions is to build equi­ travel distance between two points in a subregion. Daganzo (12) indi­ 112 travel time contours from the depot. For this scenario, the contours cates that the expected travel distance between points is k3 , where

are squares centered at the depot at 45° to the metric's preferred k is a dimensionless constant; approximately 0.82 for L1 metrics and directions. Daganzo (12) indicates that delivery subregions should 0.57 for Euclidean metrics. Thus, the local travel distance is approx- be rectangular in shape and should be oriented perpendicular to . imately lk\16, and the local travel time in a subregion in band i is: these contours, as shown in Figure 2(left). As before, the outermost delivery subregions are designed first, T3 = l;k\16 (11) followed by the subregions in bands progressively closer to the ter­ v minal. Vehicles are again bound by a time constraint that includes the line-haul time, Tl, the local travel time, 72, and the handling As in the previous section, the design process begins with the out­ time, T3. Letting D; equal the distance to the inner contour of band ermost band of the city, which faces the following time constraint: i, the line-haul travel time can be defined: Tl +T2+T3::;T (12) i Dc-2,d; Tl= D; = I (13) v v v (7) Solving the constraint gives the optimal length of the subregions in The handling time, T2, is the time required to perform the deliv­ band number 1: ery and/or pickup tasks at all points in the subregion. Assuming that the required handling time per stop, T, is constant on the average, * _ Tv - De (14) then the total handling time is the product of the total number of f, - (63) 112 TV+ kv16 - V2 points in the subregion and T. Daganzo (12) illustrates that, for an L, metric and randomly scattered points, the tour length minimizing For design purposes, and particularly for cost estimating pur­ dimensions for delivery subregions are approximately: poses, we also need to know how many delivery subregions are in each band. Having determined Li, we can calculate the average Subregion width = (6/3) 112 (8) perimeter of the band and the total number of subregions in band n:

Subregion length = C(63)- 112 (9) N = average perimeter of band i (15) ' optimal zone width where C is the number of points in the subregion. The total number of points in a subregion, then, can be estimated by solving the sec­ 4VlD - 41* 112 ]+ ond equation for C = /(63) • By so doing, the total handling time (16) for a zone in band i is approximated: N, = [ (~)•n '

T2 = T/;(63) 112 (10) where[]+ is the nearest integer greater than the quantity in brackets. Kiesling and Hansen 65

The second band, or any subsequent band, is designed in a simi­ i-1 lar process, substituting De - V22J for D, in the line-haul expres­ V2Dc - 4x - 4 I l"j - 21~ ]+ sion. Repeating this process, the following recursive design equa­ N; = (~f <+I for i~(t+ I, ... n-1), (21) tions emerge: [

Tv - De+ V2f f'J_, Li= j=2 (17) 112 (68) TV + kvl6 - V2 N; = [ (;)'." rfor i ~ n, (22)

-[ 4'\/2Dc - 8 lj_, - 4lj ]+ where j~ (18) N;- r D x= I l't+ _e_ (~t i=l 2'\/2

The above equations can be used iteratively to design each deliv­ is the distance between contour t and the edge of the city. ery subregion in the delivery area.· Note, however, that this design The third scenario (terminal in the city corner) has associated process will result in irregular delivery bands (and zones) adja­ with it a set of design equations similar in nature to the city center cent to the terminal. It may be necessary to "manually" adjust sub­ scenario originally considered: regions boundaries to cover the area in consideration, either by expanding/contracting nearby subregions or adding another subre­ gion. Whatever method is employed, the number of additional delivery zones required is small relative to the total number of zones required for the entire delivery region.

Other Terminal Locations

Air express terminals are typically located at local airports which, more often than not, are located on the perimeters of cities due to land and noise constraints. As a result, it is not always appropriate to assume that the terminal is in the city center. Two other terminal locations have been evaluated using the procedure just described; one with the terminal located in the corner of the city, and another with the terminal in the middle of the city edge. Letting De equal the FAST ROADS travel distance from the terminal to the furthest edge of the city, the Euclidean length of the city edges are le = DJV2. The models presented in the previous sections assume that vehicles When the terminal is located in the center of the city's edge, the travel the same speed on all roads. Since city networks are combi­ equi-travel time contours take on a peculiar shape. The outermost nations of local roads, arterials, and freeways, the aforementioned bands are simply formed by straight contours. But, halfway through models must be expanded to allow for more than one travel speed. the city, the contours take on a rectangular shape, as shown in Fig­ Newell (13) examines the impact that fast roads have on the ure 2(b). As a result, additional notation is required; delivery sub­ shape of equi-travel time contours, demonstrating that a single fast regions on the outermost contours are in bands 1 to (t - 1), the tran­ road stretches the contours in the direction of the road. Kiesling (14) sition band is band t, and the half diamond shaped contours form indicates that a grid of fast roads, which arguably exists in any bands (t + 1) ton. To determine the number of subregions in the major city, results in equi-travel time contours that are closely transition band, t, the zone is divided into two parts; the "cross­ approximated by a contour oriented at 45° to the origin as in the pre­ piece" (which is equivalent to bands 1 through t - 1) and the "legs" viously analyzed case in which no fast roads are present. which form the edge of the city. The number of subregions in the The approximated equi-travel time contour is dependent on pre­ cross-piece is given by Equation 20, and the number of subregions viously defined variables and two vehicle travel speeds, v (fast) and in the legs is estimated as the dividend of the area of the two legs 1 vs (slow). If we assume that delivery vehicles travel fast on the line­ and the optimal area of a subregion located in band t. Then, the haul portion of their delivery tour, and travel slow on local portions design equations can be expressed: of the tour, the original time constraint can be rewritten:

N; = [ ( i)'" ]• for i~(l, 2, ... , t- I), (19) (26)

De x Solving the constraint as in Section 2, we can determine the optimal zone lengths: · N; = [ (182 t + (1~ t Li (De + V2Dc

l* - ~Ir (27) - 2V21'; _ 4 jt, /)'_,)r for i ~ '· (20) I - (68)1/2'TVJ + k y 6v/ - '\f2 Vs 66 TRANSPORTATION RESEARCH RECORD 1506

i The final cost to include in this model is the fixed vehicle cost, Tvf- De+ Y2I r;_, j=2 Cf. The total number of vehicles required to deliver freight through­ li= (28) out the city is assumed equivalent to the number of delivery subre­ (60)112 TVJ + kv'6vf - \/2 gions in the city: Vs Fixed vehicle cost= ct(~ NII) The above expressions are true no matter where the terminal is located. However, the definition of De changes for each scenario. The total cost of delivery, TC, can then be expressed as the sum Generally speaking, De is the travel distance from the terminal to the of the aforementioned costs: furthermost point on the city boundary. The optimal zone length is now a function of two speeds. But, . TC= c(Ao + N;) + DfN; + tiN;) what about the number of subregions in each band? None of the pre­ ;~ cd(2~ k\16~ viously defined expressions change simply because, in all previ­ ously analyzed scenarios, the number of subregions in each band is + C111 (Aoz) + C{( ~ N;) (34) not a function of v. Thus, the design of all delivery subregions in metrics with fast roads is identical to the previously described To demonstrate the economic characteristics of this model, we

process with the exception that the optimal zone length changes. consider a diamond-shaped city with an L 1 metric and a terminal located in the lower corner. In such a case the city area, A, is l?, or Dl!2. Table I summarizes the parameter values assumed for the LOGISTIC COSTS OF GROUND-SIDE DELIVERY remainder of this section. The results of the subregion design algorithm confirm a priori The heuristic developed and demonstrated up to this point provides expectations about the subregion partitioning, that bands furthest all of the information needed to locate and size delivery subregions from the termi~al are narrowest (l l.74 km), and bands closest which, in turn, allows us to begin analyzing ground-side pickup and 1 = to the terminal are widest (l 6.13 km) with the exception of band delivery costs. In this section, a basic cost model is developed that 7 = n. It is easily shown that bands i ton i increases by the constant incorporates four categories of logistics cost; costs of stopping, + percentage costs of overcoming distance, costs of carrying additional freight, and fixed vehicle costs. Included in the time constraint that underpinned the development of our design algorithm is the time required to stop at an origin or destination and move the package to/from the delivery vehicle. Sev­ eral costs are incurred each time the delivery vehicle stops includ­ To assess the concepts of scale and density economies in ground­ ing labor, vehicle depreciation costs and materials. The total cost of side delivery operations, the total cost formulation is used to deter­ stopping is determined by assuming that the cost per stop, C.,, is con­ mine average costs (total cost per package) of delivery under vari­ stant on the average. Including the entire ground-side delivery sys­ ous assumptions. Scale economies are defined as a change in the tem, the total number of stops is the sum of city-wide stops and average unit costs of production resulting from a change in output. stops at the local terminal. Thus, the total stopping cost follows: If output is defined as the number of points visited by a carrier, which increases as a result of city growth (De increasing, ceteris paribus) or an expansion in the carriers delivery market, it is easily Stopping cost = + (29) Cs (Ao ~ N;) shown that there are diseconomies of scale. If o increases while holding De and all other variables constant (which is more accu­ where A is the area of the region in question. rately called economies of density), it is clear that there are pro­ The cost per mile, Cd, is also assumed constant on the average. found economies of density in ground side freight distribution, as Recalling the need to include both line-haul trips in the cost formu­ indicated by the decreasing average unit cost curve in Figure 3. lation, the total distance traveled is easily determined: Although the finding of such significant economies of density is not

Total distance= 2f DiNi + kV6f tiN; (30) i=l i=l TABLE 1 Assumed Parameter Values for Demonstration of Cost /1 11 ) Model Distance cost= Cd ( 2i~ DiN; + k\16~ tiN; (31) Parameter Assumed Values i 0.39 to 9.67 stops/km2 (1 to.25 stops/mi2) where DT= De - It1 and D~ = 0. j=I 40.2 km (25 mi) 1 to 4.5 hr We also include the added cost per item carried, C,n, in the for­ 64.4 kph (40 mph) 32.2 kph (20 mph) mulation. The total number of items carried is the product of the 5 min number of stops and the number of packages per stop, z: 0.82 1 to 4.5 pkg/stop

Marginal cost = C111 (Aoz) (32) $0.62/km ($I/mi) $2/stop The marginal costs are very small compared to other logistics costs $0.05/pkg and are frequently ignored. $1.25/veh Kiesling and Hansen 67

$2.50

$2.45 •

$2.40

C1> 0) ca • ~ $2.35 u ca Q.. • • ~ $2.30 Q. • • u; • • 0 $2.25 • 0 • • • • • • • • • • • • • $2.20 • • •

$2.15

$2.10

0 2500 5000 7500 1 0000 12~00 15000 17500 20000 22500 25000 Number of Packages

FIGURE 3 Economies of density in ground-side pickup and delivery operations.

surprising, this result has particular importance in the analysis of The trade-off between economies of density and "economies carrier competition when coupled with the impact that the time con­ of time" can be illustrated two ways. First, we can assess the impact straint has on operational costs, discussed below. of varying Ton average unit costs, taking into account the decrease As discussed in the introductory section, time may be the most in available demand caused by an increase in T. Clearly, at T = 0, important strategic decision variable facing air freight carriers. the maximum number of points are potentially served (although There are two competing effects of varying the time constraint, T. there is no way to service the pickup and delivery points in zero First, increasing T lowers average unit costs significantly, ceteris time). Furthermore, it is appropriate to assume that the entire mar­ paribus. The reason is simply that if T increases, the number of ket (oA) is available for time constraints up to 2 hr. For T greater required delivery subregions decreases, thus lowering the opera­ than approximately 2 hr, however, the number of points that can be tional costs. In a competitive setting, it may appear that a combina­ serviced begins to decline for the previously discussed reasons. tion carrier would therefore want to make the time constraint as Eventually, no demand is available at T = 9 hr, the full business large as possible to mi_nimize costs. There are several potential day. The available demand (AD) distribution could be represented trade-offs to increasing T, however. To illustrate, consider the ways as follows: in which an increase in T can be accomplished. First, it can be accomplished by setting the pickup deadline earlier (or the delivery AD= o(l -( eT-y )x) (35) deadline later). However, this would diminish the level of service } + eT-y offered to the customer, causing some customers to select another carrier or not purchase the product at all. Second, it could also be where x and y are distribution shape parameters, assumed to be accomplished by serving a smaller market (within a city). In other 2 and 4 for demonstration purposes. The available demand (as a words, reduce the area served from a terminal. Third, it could be function T) is illustrated in Figure 4. Substituting this "available accomplished by reducing the number of destination cities served demand" quantity into the total cost model (Equation 34) and vary­ from an airport. A combination carrier, for example, may have ing T from 0 to 9, Figure 5 illustrates that the cost minimizing departing flights to Phoenix and Chicago at 6:30 p.m. and 7:00 p.m., time constraint is from 5 to 7 hr, but the improvement in costs over respectively. With a 4:30 package pickup deadline, both destina­ T = 3 is relatively small. Profit maximization is more impor­ tions are served by a 1.5-hr time constraint (allowing 0.5 hr to load tant than cost minimization for a freight carrier, however, so a aircraft). T could be increased to two hours if only Chicago bound more appropriate way to view the effect of time on production packages are served. Thus, T can be increased in several ways, strategy is to consider the impact that the time constraint has on including combinations of the above methods. The trade-off, how­ carrier profits. Assuming an average price per package of $10, Fig­ ever, is that any of the aforementioned "solutions" reduces the ure 6 illustrates that the profits of a hypothetical carrier are maxi­ demand served by the carrier, which increases average units costs mized when the time constraint is approximately 1.5 hr (for this according to the previously illustrated economies of density. example). 25 :!' • E cT • ti) 20 • ._ Q) • Q.

ti) • c 1 5 ·5- -Q. • ~ • ti) c 1 0 Q) • c ._ • Q) E 0 5 • ti • :::J () • • • • 0 • • • 0 2 3 4 5 6 7 8 9 Time Constraint (hrs)

FIGURE 4 Available customer density as a function of time constraint, T.

$3.50 • $3.00

$2.50 Q) C> CV • ~ 0 • CV $2.00 a.. • Ci> Q. $1.50 a; • • 0 • () • $1.00 • • • • • • • •• • • • • • • • • • $0.50

$0.00 0 2 3 4 5 6 7 8 9 Time Constraint (hrs)

FIGURE 5 Average unit costs as a function of time constraint, T. Kiesling and Hansen 69

160 • • • • 140 • 120 • -0 0 100 0 • 0 80 • -....0 [l. 60 • • 40 • • 20 • • • • 0 • • • 0 2 3 4 5 6 7 8 9 Time Constraint (hrs)

FIGURE 6 Carrier profit as a function of time constraint, T.

CONCLUSIONS REFERENCES

Throughout this report, several aspects of the design of ground-side 1. Sigafoos, R. A. Absolutely Positively Ovemight: The Unofficial Corpo­ delivery systems have been explored. It was first indicated that the rate History of Federal Express. St. Lukes Press, Memphis, Tenn., 1988. binding design constraint is that of time, not vehicle capacity con­ 2. Trimble, V. H. Overnight Success-Federal Express and Frederick Smith, Its Renegade Creator. Crown Publishers, Inc., New York, 1993. straints as assumed by most previous studies. It was further indi­ 3. Shaw, K. Airlines Join RPS to Compete with Express Carriers. Air cated that an appropriate way to approach the design of delivery Cargo World, Vol. 82, No. 4, April 1992, p. 40. subregions is to first define the time constraint as a function of line­ 4. Kiesling, M. K., and M. M. Hansen. Integrated Air Freight Cost Struc­ haul time, handling time, and local travel time. Then, according to ture: The Case of Federal Express. Institute of Transportation Studies, UCB-ITS-RR-93-15. University of California, Berkeley, 1993. this time constraint, design the delivery subregions along the equi­ 5. Daganzo, C. F. The Length of Tours in Zones of Different Shapes. travel time contours beginning in the outermost band and iteratively Transportation Research, Vol. 18, No. 2, 1984, pp. 135-145. moving toward the terminal. Expressions were derived for the sub­ 6. Daganzo, C. F. The Distance traveled to Visit N Points with a Maximum region dimensions, the number of subregions per band, and the of C Stops per Vehicle: An Analytical Model and an Application. location of the subregions for cities employing Li metrics with ter­ Transportation Science, Vol. 18, No. 4, November 1984, pp. 331-350. 7. Newell, G. F., and C. F. Daganzo. Design of Multiple-Vehicle Delivery minals located centrally, in one corner of the city, and in the middle Tours. I. A Ring-Radial Network. Transportation Research, Vol. 20, of one edge of the city. No~ 5, 1986, pp. 345-363. City street networks generally allow for more than one speed of 8. Newell, G. F., and C. F. Daganzo. Design of Multiple-Vehicle Delivery travel. Fast roads, as they are often called, significantly change the Tours. II. Other Metrics. Transportation Research, Vol. 20, No. 5, 1986, pp. 365-376. shape of the equi-travel time contours that the design is based on. As 9. Newell, G. F. Design of Multiple-Vehicle Delivery Tours. III. Valuable a result, the impact of fast roads in a city network was explored. The Goods. Transportation Research, Vol. 20, No. 5, 1986, pp. 377-390. design framework was then generalized to allow for two travel 10. Langevin, A., and F. Soumis. Design of Multiple Vehicle Delivery speeds; fast travel on line-haul trips and slower travel on local streets. Tours Satisfying Time Constraints. Transportation Research, Vol. 23, No. 2, 1989,pp. 123-138. The results of the design process for a square city with an Li met­ 11. Han, A. F.-W. One-to-Many Distribution of Nonstorable Items: ric, two travel speeds, and a terminal in one corner, were then used to Approximate Analytical Models. Ph.D. dissertation. University of Cal­ develop a total cost model of ground-side pickup and delivery opera­ ifornia at Berkeley, 1984. tions. The model, in turn, was used to explore the economic cost struc­ 12. Daganzo, C. F. Logistics Systems Analysis. Springer-Verlag, New York, 1991. ture of the delivery system. It was indicated that ground-side pickup 13. Newell, G. F. Traffic Flow on Transportation Networks. The MIT and delivery operations exhibit significant economies of scale and Press, Cambridge, Mass., 1980. profound economies of density. The results were highly robust with 14. Kiesling, M. K. The Modelling of Logistic Costs in the Air Freight respect to changes in all design variables. It was also demonstrated Industry. Ph.D. dissertation. University of California at Berkeley, 1995 that as the time constraint increases, the average unit cost decreases. (forthcoming). However, increasing this decision variable results in a decrease in the Publication ofthis report sponsored by Committee on Airport Terminals and market that is potentially captured by a competing carrier. Ground Access. 70 TRANSPORTATION RESEARCH RECORD 1506

Challenges in Developing an Airport Employee Commute Program: Case Study of Boston Logan International Airport

DIANE M. RICARD

Boston Logan International Airport, a major trip generator, contributes port has been successfully promoting, maintaining, and improving to and is impacted upon by traffic congestion in the Greater Boston area. an aggressive air passenger ground-access program aimed at reduc­ Located about 3 km from downtown Boston, Logan is the fifth largest ing the number of vehicle trips per air passenger trip. The nature of airport in the United States in terms of origin-destination air passengers. the Logan working environment has made it much more difficult to Logan origin-destination passengers begin or end their air travel in the Boston region and affect the Boston regional transportation system. provide a comparable access program for employees. Because air passenger growth must be accommodated within the exist­ Within the existing menu of transportation services available to ing airport boundaries and regional roadways, restrictions imposed by air passengers and downtown commuters, Massport offers some the Logan Airport Parking Freeze, and by a responsibility to help reduce incentives to employees who are seeking alternatives to the drive­ regional environmental impacts, it has become increasingly important alone commute. However, many Logan employees cannot com­ to find feasible ways to reduce the vehicle trip generation rates of the mute using these services, because they were developed for cus­ various Logan Airport user groups. The commuting patterns of the tomers with different travel requirements. 16,000 Logan employees, who account for about 20 percent of average annual weekday traffic, are characterized in this paper. Data presented In this paper the challenges Massport faces in developing a suc­ in the paper are based on the results of an airport employee survey con­ cessful, cost-effective employee access program are presented ducted in 1990. Commute profiles of both flight crews (who exhibit through a description of the access characteristics of Logan and travel characteristics similar to those of air passengers) and non-flight the Logan work environment. The commuting patterns of Logan crew employees are highlighted. Since the airport is staffed 24 hours a employees are characterized, available commute options are de­ day with various types of workers, feasible solutions to reduce airport scribed, and initiatives that Massport currently offers or is consid­ employee trips will be different from measures tailored to influence commute habits of the traditional office workers. In this paper available ering in the future to encourage commuting in higher-occupancy alternatives to the single-occupant private automobile are discussed, modes are discussed. and their effectiveness relative to employee demand is assessed. There is a small employee market for most alternatives currently available, as each service was developed primarily for air passengers or downtown commuters. Finally, a summary is presented of the initiatives that the ACCESS CHARACTERISTICS OF LOGAN Massachusetts Port Authority (owner and operator of Logan Interna­ AIRPORT (1) tional Airport) has taken in the past and is considering in the future to encourage employees to use higher-occupancy commute modes. A number of characteristics of the Boston regional transportation system, the placement and size of the airport, and the airport's prox­ Boston Logan International Airport, owned and operated by the imity to neighborhoods all make access to Logan difficult during Massachusetts Port Authority (Massport) is a major trip generator. certain times of the day and week, and present significant operations Logan contributes to and is affected by traffic congestion in the management challenges for Massport. The characteristics may be greater Boston area. Located about 3 km from downtown Boston, categorized as follows: Logan is the fifth largest airport in the United States in terms of ori­ gin-destination air passengers. Logan origin-destination passengers 1. The greater Boston area and the New England region are prin­ begin or end their air travel in the Boston region, and affect the cipal destinations for both business and pleasure travelers, and are Boston regional transportation system. located on one of the most heavily-traveled air corridors in the U.S. Reducing air passenger and employee vehicle trips is important (Boston, New York, and Washington, D.C.). Ninety percent of to Massport from an air quality perspective and from an airport man­ these passengers are using the Boston local transportation infra­ agement perspective. On an average weekday in 1992, about 85,000 structure to access Logan (rather than flying into and out of Logan vehicle trips were generated by Logan Airport. By comparison, on their way to another destination). Boston proper generated about 756,000 vehicle trips on an average 2. The airport is served by a limited number of access routes (two weekday in 1992 (source: Central Transportation Planning Staff, cross-harbor tunnels and one bridge), which do not haye the capac­ 1992 Interim Regional Model Set). About 60 percent of vehicle trips ity to handle easily the volume of traffic using the system during to and from Logan are made by air passengers, and another 25 per­ periods of high demand. cent are made by Logan employees. For a number of years, Mass- 3. Because of Logan's proximity to downtown Boston, the regional highway and public transportation system is focused on Assistant Director of Economic Analysis, Massachusetts Port Authority, 10 Boston in a series of radial routes converging on the central busi­ Park Plaza, Boston, Mass. 02116-3971. ness district. Airport traffic mixes with regional vehicular traffic, Ricard 71 and use of the subway, commuter rail, and bus system requires employee commute survey administered in the spring of 1990, transfers in downtown Boston to reach the airport. The length of which was answered by 15 percent of employees. Since the Logan time required to reach the airport and the inconvenience of the trans­ work environment and available commute options have not fers makes public transportation a difficult choice for frequent air­ changed much since 1990, Massport believes that the survey con­ port users, such as employees. tinues to explain overall commuting behavior. A copy of the survey 4. Due to the restricted availability of land for airport develop­ instrument is included as Figure la. Table 1 is a summary of perti­ ment and Massport' s commitments to the neighborhoods around the nent commute characteristics of Logan Airport employees. airport, the size of Logan Airport is fixed at its present area. Mass­ A separate survey form was administered to Boston-based flight port is committed to accommodating airport growth within the crews. All of the survey questions were the same as in the non-flight existing airport boundaries. crew survey, with the exception of the following replacements (Fig­ 5. Responding to environmental and community responsibili­ ure 1b ). Question 22 on the non-flight crew survey was not asked ties, Massport has committed to a moratorium on the number of air­ of flight crews. port parking spaces and to limiting traffic accessing the airport by way of local neighborhood streets. As part of this moratorium, Commute Modes Mass port has committed to relocating about 30 percent of on-airport employee parking spaces to off-airport locations. Currently there are time and cost incentives for most Logan employees to commute by automobile compared to alternative THE LOGAN WORK ENVIRONMENT modes. As at most U.S. airports, the majority of Logan employees enjoy parking privileges on or near the airport, fully subsidized by The Logan work environment presents a challenge for developing their respective employers. All of the major airlines subsidize a responsive and cost-effective employee commute program. Stan­ employee parking at airports throughout the United States. It would dard transportation demand management options, such as flextime, be difficult for an airline to discontinue this benefit at some airports vanpooling, or carpooling on a regular basis, are an option for only and provide it at others. In terms of competitive hiring, it would also a small proportion of the population due to nontraditional work be difficult for an airline to discontinue the employee parking sub­ schedules. sidy unless other airlines were doing the same thing. Logan Airport operates 365 days a year, 24 hours a day. Holidays Almost all of the surveyed employees reported that their and popular vacation periods, when many businesses slow down, employer does not subsidize public transportation. On an average are some of the busiest times at an airport. Because the almost weekday about ~O percent of Logan employees commute to Logan 16,000 people are employed at Logan to maintain its continuous by automobile, and most of them are commuting alone. Another 10 operation, the concentration of employees commuting during s.tan­ percent of employees commute by subway, and the remainder walk, dard business hours is sparser than in other industries. On an aver­ take a bus, or use other means to get to Logan. By comparison, the age weekday, only 60 percent of all employees commute to Logan, transit share of employees commuting to Boston proper is 44 per­ and only 25 percent of all employees arrive between 6:00 a.m. and cent (source: Central Transportation Planning staff, based on the 10:00 a.m. Between 30 percent and 40 percent of employees staff 1990 Census). the airport on Saturdays and Sundays. Of employees commuting by subway, less than half have access There are 140 employers at Logan. Seven major airlines are to employer-subsidized parking. Twice the proportion of automo­ responsible for about 55 percent of employees, and Massport bile commuters have access to employer-subsidized parking as do employs about 4 percent. Many employee work schedules are subway commuters. Subway commuters have fewer automobiles related to air passenger demand and flight operations. These per adult available in the household compared to automobile com­ employees may be subject to either scheduled or nonscheduled muters. The data indicate that the majority of employees commut­ overtime, and do not have flexibility in their work schedule. Non­ ing by subway are doing so because an automobile or airport park­ scheduled overtime is tied to flight delays and cancellations, events ing is not available to them. that are very difficult to predict and plan for. The numerous airport­ From most towns, depending on the time of day, the commute wide shifts, which vary by company, make it difficult to develop time by automobile offers the employee a noticeable time sav­ commute options around particular shifts. ings over scheduled high-occupancy vehicle services. This will be Many Logan employees have benefits packages that are based on discussed further under the section Employee Alternatives to the contractual agreements or national company policies. Groups of Automobile. workers at Logan belong to various unions. A collective bargaining agreement may limit the incentives to high-occupancy commuting Geographic Concentrations or disincentives to the drive-alone commute that can be offered to a group of employees. Similarly, airline-wide employee policies may The 10 towns with the largest concentrations of Logan employees limit what a local airline station manager may provide employees are all in the immediate vicinity of the airport. For the 40 percent of as alternative commute incentives, especially now, when airlines Logan employees residing in these towns, Logan is a quick trip by are trying to cut costs to remain competitive. automobile. For many, the extra time involved in carpooling or using public transportation is viewed as an unnecessary inconvenience. LOGAN EMPLOYEES Job Categories The following is a brief profile of Logan Airport employees, pre­ sented to show the uniqueness and complexity of an employee pop­ Twenty-five percent of Logan employees are traditional office ulation at a major airport. Most of the information is based on an workers, 25 percent hold sales or service related positions, and 25 Logan Airport

Employee Survey ::::::::::::::·:·:·:·:·:.:.·.:·:-:-:::-:-:::::-:-:-·-·.··· ············· ...... Why you have been given or sent this questionnaire

This survey is being carried out to give Massport an up-to-date picture of the travel needs of people who work at the airport. To plan for the Third Harbor Tunnel and 3. other developments, we need to find out how airport The d~y on which you arrived at Logan on that employees are currently getting to and from work. occasion was ... (check one day only)

Please take a minute or two to answer these questions. 1 0 Monday 5 0 Friday 2 0 Tuesday , D Saturday If you received this questionnaire at home ... 3 0 Wednesday 1 O Sunday , 0 Thursday please fill it out today, fold it so that the return address shows on the outside, then put it in the mail. 4. On that day, from where did you start your trip to go to work at the airport? If you received this questionnaire at work ... Please specify: please fill it out today and drop it in one of the marked boxes at your workplace; or fold it so that the return address shows on the outside and put it in the mail. town or city state zip code Don't return a questionnaire at your workplace if you've already returned one that you got in the mail. street address or nearest intersection Thank you for your help; it is important to us. All replies office use only: are confidential.

5. Is this place... (check one only) Sincerely, , 0 your own home? 2 D some other place?

Patrick B. Moscaritolo 6. At what time did you get to the airport on that day? Director, Logan Airport hr. min. Entertime:ll: If, 0A.M. LLJ LLJ 2 D P.M.

1. Today's date is ... (Check one day and fill in date) ,O Monday aD Friday 7. How did you arrive at the airport on that day? 20 Tuesday ,D Saturday (Check one only) 3 D Wednesday 1 D Sunday __/ __ /90 1 D driving a private vehicle (car, van, or light truck) .. D Thursday month date 2 D passenger in a private vehicle

2. Think back over the last seven days. On which of those days did you go to work at Logan? (Check every day on which you traveled to work at the airport) ,O Monday aD Friday 20 Tuesday eD Saturday 3 D Wednesday 10 Sunday .. O Thursday

FIGURE la Logan Airport employee survey. Ricard 73

8. If you went to work by private vehicle on the day 14. After work on that occasion, at what time did you In question, how many people were in the vehicle leave the airport? In total, Including yourself? r-1 hr. min. Enter number: LLJ Enter t.ime: 10A.M. 2QP.M.

9. Was the private vehicle parked ... (check one only) 15. You've now told us how you traveled to and from 1 D in your employer's parking area immediately your work at the airport on one particular occasion. beside your place of work? How do you usually get to and from work? Is that 2 O in Massport's Bird Island Flats employee lot? in the same way as on the occasion you've told us about? (Check one only) 3 O in a passenger parking garage or lot (Terminal A, B,D, or E, or Central Parking)? 1 Dyes, I usually go to work in the way I've ~ D near the airport, off airport property? already described ·

5 O or was it driven away from the airport, and not 2 D no, I usually go by car or van that is parked a1 parked nearby? the airport e 0 other (spedfy) 3 0 no, I'm usually dropped off (and picked up) at the airport by a car or van • 0 no, I usualty go by public transportation (or walk) 1o. If the private vehicle was parked at or near the 5 D other (specify) airport, what was the cost to you for parking? If some of the costs are paid by your employer or by other passengers, count only your own share of the costs.

1 D parking didn't cost me anything 16. If for some reason the transportation you usually use to get to the airport wasn't possible for a long period of time, how would you get to work? 20 parking j per day $' II 10 (Check one only) cost me _.__ ...... --£..__,.LLJ 2 0 per month , D by private vehicle (car, van, light truck) parked 11. On your private vehicle trip to work, did you make at the airport any stops to pick up or drop off other people, or to do errands? 2 D dropped off (or picked up) by private vehicle (car, · van, light truck) 1 0 yes, I made stops on the way 3 0 MBTA Blue Line 2 D no, I came straight to the airport • D Logan Express bus 12. On your private vehicle trip to work, did you s 0 other bus service travel... (check all that apply) 6 0 other means (sf>6Cify) ------

1 D through the Callahan Tunnel? 2 0 on the Central Artery, from the south C-South Station tunner)? ,O on the Central Artery, from the north? "0 on Route 1, from .the north? 17. Which of the following statements best applies to 5 0 by none of these routes? you? (Check one only) ·

13. On the day In question, did you leave the airport , 0 I have onty one job at the airport again before you finished work? 2 Q I have two or more different jobs at the airport (Check all answers that apply to you) L How many different jobs do you currently have? 1 0 no, I stayed at the airport until I finished work 2 0 yes, my job requires me to spend most of my Enter number: D working time away from the airport , D yes, I left by private vehide (car, van, or truck) 18. For which company do you work at the airport? "0 yes, I left by other means (MBTA. ferry, taxi, (If you have more than one job, give the one you walk, etc.) spent most time at on the occasion you described If you did leave the airport during your working earlier) period, how many times did you leave the airport for... Enter company name:. ______... job-related purposes? office use only:I._ _.____.____, ... personal reasons?

FIGURE lb Logan Airport flight crew survey questions. percent are flight crew members. The remainder hold a variety of occupancy vehicle alternatives potentially available to an individ­ positions, including maintenance and ramp service. Only about 15 ual employee. percent of employees have more than 15 min of flexibility in their Massport, through survey analysis, has found no significant dif­ work schedules. This suggests that of the employees holding posi­ ferences in commute patterns of employees by job classification, tions that are not firmly fixed to a schedule, most are already eligi­ with the exception of flight crew members compared to non-flight ble for flextime. Flextime increases the number of carpool and high- crew members. 74 TRANSPORTATION RESEARCH RECORD 1506

19. Where at the airport do you report to that job? (Check one only) , 0 Terminals, old or new towers, or parking garages 2 D north cargo area (Marriott kitchen, Delta hangar, NW!TWA hangar, Pan Am freight, etc.) 3 0 south cargo area (Bird Island Flats, Mass Tech Center, general aviation, pos1 office, etc.) , 0 rental car facilities 23. Unless you already did so at question 4, please s 0 Hilton hotel, MPA heating plant, or tell us where your home Is: American air freight 6 0 other (specify)

city()( toYm Stale zip code

20. How much fle·xibility do you usually have in choosing the time you report for work? (Check one only) street address or nearest inte

24. How many of the people living in your household 21. Sometimes employers share some of their (including you) are ... employees' costs of getting to and from work, or of parking at work. In the following list, please check any form of help that you personally qualify for(even if you don't use It). . .. aged 17 or over Enter number: (Check all that apply) [lJ , D employer provides free parking spaces 2 D employer pays part (or all) of the cost of parking ... aged 16 or less Enter number: [lJ 3 0 employer pays part (or all) of the cost for public transportation , D employer provides car (or helps with car purchase ... employed for 1 O or loan) or more hours s D employer reimburses gasoline or mileage costs per week? Enter number: [lJ , D employer reimburses tunnel or highway tolls 1 D none of the above

25. How many private motor vehicles (cars, vans, or 22. Which one of the following categories best light trucks) are available for use by members of describes your job at Logan? (Check one only) your household?

1 D executive/managerial 2 D professional!tectinical Number of vehicles available: [lJ , D administrative support/clerical , D sales/service s D production/crafts s 0 maintenance/trade

1 D other (specify) ------

FIGURE 1 (continued)

Flight Crew Members (called a tour of duty). A tour of duty will begin and end at Logan Airport, but often it lasts for several days. The average tour of duty As mentioned above, about 55 percent of employees work for seven for Boston-based flight crew members is 3 days, meaning the egress major airlines at Logan. But the group of all airline employees is not trip from Logan is taken 2 days after the access trip. necessarily an easy target for trip reduction, since 45 percent of this Due to employee flying time restrictions imposed by the Federal group are flight crew members, and a good deal of them do not Aviation Administration and individual airlines, a flight crew mem­ travel during peak commute periods. Flight crews, accounting for ber does not generate many commute trips per month. For instance, 25 percent of Logan employees, are only responsible for about 10 Federal Aviation Regulations prohibit flight crew members from percent of average weekday commute trips. flying more than 1,000 hours per year, or 100 hours per month, or Flight crews are the pilots and flight attendants who are based in from flying for more than 30 hours in any seven-day period (2). For Boston and commute to Logan to begin their flight assignment most flight crew members, a private automobile is used to commute Ricard 75

26. Employee parking spaces are likely to be less conveniently located during the construction of the Third Harbor Tunnel. In your opinion, how can Massport help make your commuting more convenient?

office use only: ~

Completely optional:

If you would like to enter our drawing for five free dining-out certificates, we need your name and telephone number. Otherwi58, you may leave this blank.

Your name: please print

Telephone:~~~-'-~~~~~~---~~~~~~~ area code

This number is 1 work 2 home 0 D

FIGURE 1 (continued)

to Logan, and is parked at Logan for the duration of the tour of duty. in developing reasonable commute options for this employee The irregularity of commute hours and days make it difficult for a group. flight crew member to participate in a carpool or vanpool. It is par­ ticularly difficult for a flight crew member to plan on using an alter­ native to the single-occupant automobile for the trip from Logan to Non-Flight Crew Members home, since the timing depends on the arrival time of a scheduled flight, which may experience delays. Furthermore, flex-time is not Non-flight crew employees are responsible for the continuous oper­ a consideration for a flight crew member, as work assignments are ation of Logan Airport. On an average weekday, 90 percent of com­ scheduled around specific flights. mute trips are made by this catch-all group of employees. A Pilots tend to live farther away from Logan compared to non-flight crew member generally has a work day that is be­ other Logan employees as a result of higher-than-average incomes tween eight and ten hours. Individual work schedules take on a and the need for fewer average commute trips per month. Their range of forms including standard business hours, fixed shifts, and sparser geographic concentrations further accentuate the difficulty varying hours by day, by quarter, or other time increments. Some 76 TRANSPORTATION RESEARCH RECORD 1506

13. For your tour of duty ("trip" or "block") that began with the trip to work that you have just described, on what day of the week did (or wllij you end back at Logan airport and leave work?

,O Monday go Friday 20 Tuesday .O Saturday ~ D Woonesday 10 Sunday ,O Thursday

14. On that day, at what time did you (or do you expect to) leave the airport after finishing work1 hr. min. Enter f.Jme:. DUJ·D.A..M.: I O 2 P.M.

17. Are you... (Check one only) , D a flight attondant

2 0 a cockpit crew member

18. For which airline do you work?

Enter airline name: ------::::::======:::::: offiC6 use only:!~_.._____.___.

20. In what year did you first start to work from a base at Logan Airport, in either your current or a previous job? Enter year: 19 Li]

FIGURE 1 (continued)

are not required to be at the airport at particular fixed hours, others next section describes services currently available and their ability have scheduled overtime, and others are subject to nonscheduled to accommodate Logan employees. overtime.

EMPLOYEE ALTERNATIVES TO THE Logan Employee Potential for Traditional Carpooling AUTOMOBILE

The variety of work schedules and inflexibility of work hours limit A variety of high-occupancy modes serve the airport, including sub­ the pool of employees that can take advantage of traditional car­ way, ferry, limousine and publicly and privately operated scheduled pools or vanpools as an alternative to the drive-alone commute. bus service. All of the services were developed primarily for air pas­ Flight crew members have an uncertainty in timing the egress trip sengers or downtown commuters. that may be better served by an on-demand service or a regularly Keeping in mind that the majority of Logan employees have scheduled high-frequency service. This· is also the case for access to an employer-subsidized parking space at Logan, for one non-flight crew employees subject to unscheduled overtime. The or several of the following reasons the high-occupancy services Ricard 77

TABLE 1 Select Logan Airport Employee Commute Characteristics

available to Logan employees do not compete very well with the Travel Time automobile, and collectively attract only about 12 percent of Logan employees. The door-to-door commute time to Logan on some scheduled ser­ vices is not competitive with the automobile, due to the transfers required to complete the trip or to multiple stops and layovers built Geography into a trip.

The concentration of employee origins is very different from the Multiple Stops and Layovers concentration of air passenger origins; more than 50 percent of employees originate from the corridor immediately north of Logan The primary market for most of the privately operated bus routes is Airport, compared to 10 percent of air passengers. About 45 percent the downtown commuter. Offering the additional 2-mi trip to Logan of air passengers begin their trip to the airport either from Boston or is a low-cost method of filling otherwise empty seats. As such, the corridor west of Boston, compared to 10 percent of employees. travel times are minimized for commuters, and airport passengers To further accentuate the differences in passenger and employee experience longer travel times. On the way to Logan, commuters origin densities, the ratio of passengers to employees traveling to are first discharged at one or two locations in downtown Boston. Logan on an average weekday is about three to one. Scheduled bus Depending on traffic levels, travel from Logan to downtown Boston and limousine routes serving Logan were developed based on air can take from I 0 min to an hour. Because of the uncertainty of the passenger origins, and do not serve a large employee market. Table travel time between Logan and downtown and the necessity to meet 2 is a comparison of air passenger and employee concentrations by the evening schedule for commuters, the layover in Boston may be geographic zones, and Figure 2 is a map denoting the zones. as long as an hour for an airport user departing Logan. On the way into Boston, a private bus route may stop in several towns to pick up passengers. Multiple stops and potentially long layovers in Hours of Operation downtown Boston may be acceptable to the occasional air traveler, but are not acceptable service characteristics for Logan employees. For Logan employees with at least one trip end outside normal busi­ ness hours, schedules developed around air passenger or commuter peaks often do not offer the frequency of service or hours of oper­ Transfers ation necessary to provide a reasonable commute. Figure 3 is a comparison of Logan employee arrival and departure times on an The subway system Massachusetts Bay Transportation Authority average weekday. (MBT A) in the greater Boston area offers low-cost, frequent, con- 78 TRANSPORTATION RESEARCH RECORD 1506

TABLE 2 Comparison of Logan Airport Employee and Air Passenger Origins by Zone

1 Boston 4% 21% 2 Inner Rin , North 47% 5% 3 Inner Rin , Northwest 5% 4% 4 Inner Rin , West Northwest 1% 8% 5 Inner Rin , West 3% 7% 6 Inner Ring, South 8% 3% 7 Outer Rin , North 8% 4% 8 Outer Ring, Northwest 5% 6% 9 Outer Rin , West Northwest 1% 4% 10 Outer Rin , West 1% 8% 11 Outer Rin , Southwest 1% ·33 12 Outer Rin , South 6% 6% 13 Other Massachusetts 4% 8% 14 Connecticut, Rhode Island 1% 2% 15 Maine, New Hampshire, Vermont 5% 9% 16 Rest of the World 0% 2% Source: 1990 Employee Survey, Spring 1993 Air Passenger Survey. Notes: 1. Represents employee and air passenger origins on the average weekday. 2. On an average weekday, there are about 27,000 air passengers traveling to Logan and about 9,600 employees traveling to Logan; ie, 3% employees is equivalent to 1% air passengers. venient service to downtown travelers. It is connected to a network subway user. Commuter rail users must make three transfers. The of bus routes and commuter rail lines. The routes are primarily more transfers people are faced with, the less likely they are to use geared to radial travel into downtown Boston, and travel becomes a service. less convenient between points outside of downtown Boston, including Logan. There is only one direct subway line to Logan. The airport MBT A station is located on the edge of Logan Air­ Fares port, about 1 mi from the air passenger terminals. Massport offers free bus service between the station and the air passenger termi­ The fares of many of the privately operated high-occupancy vehi­ nals. Although the bus headways are consistent with the subway cle services are too high for airport employees who commute on a schedule, this presents an additional transfer for subway users. The regular basis. Generally the monthly commuting expense for the number of transfers varies between one and three for the airport privately operated bus and limousine routes to Logan is consider-

:- - i- - 1 I 1 I 1 I_ ~.L-.J. _ __L.-L---T

FIGURE 2 Geographic zones, eastern Massachusetts. Ricard 79

20 ------

15

c ~ 10 a..cu

5

0 ~~ ..... 12a.m. 2a.m. 4a.m. 6a.m. sa.m. 10a.m. 12p.m. 2p.m. 4p.m. 6p.m. 8p.m. 10p.m. Hour Beginning

• Employee Arrivals --0- Employee Departures

FIGURE 3 Logan Airport employee arrivals and departures by hour.

ably higher than the equivalent cost of services operated by the To further encourage employee ridership on Logan Express, regional transit authority. beginning in January 1995, a 4:30 a.m. bus will be added on all routes to accommodate employees with early morning shift start times. Massport estimates that additional employee pass sales will EMPLOYEE HIGH-OCCUPANCY VEHICLE cover about 40 percent of the incremental cost of the trip, and that INCENTIVES additional air passengers using the service will cover another 50 percent of the cost. Massport has been responsible for successful initiatives which The Airport Water Shuttle, a ferry service between Boston and have resulted in employees choosing alternatives to the private Logan, offers a 60 percent discount for all Logan employees when automobile. The following are the initial elements of our employee tickets are purchased in 10-ride booklets. Some of the private high­ commute options program. occupancy vehicle services offer slight discounts to regular users or Logan employees. With the exception of those offered by a couple Incentives for All Airport Employees of private operators, the discounts are not deep enough to influence employee travel behavior. Logan Express is a direct, non-stop bus service operated by Mass­ As part of an effort to reduce employee vehicle trips through local port. The three routes serve remote locations, one about 32 kilome­ neighborhoods, to comply with a federal and state regulation to ters west of the airport, one about 19 kilometers south of the airport, reduce on-airport employee parking, and to increase air passenger and one about 24 kilometers north of the airport. The services are parking, in August 1994 Massport relocated about 1500 employee operated daily, with weekday service at 30-min intervals from 5:00 parking spaces to a new parking garage one town west of the air­ a.m. until midnight. port. A bus service transports employees between the airport and the Massport offers a discounted monthly Logan Express pass for all garage. Massport estimates that some employees will switch to Logan Airport employees. The pass is priced slightly lower than the alternative modes of access rather than experience the inconve­ monthly rate for employee parking and is equivalent to between nience of driving to the remote garage to be bused to the airport. In seven and nine one-way trips on the Logan Express. Ten-ride dis­ fact, the airlines that are preparing to subsidize Logan Express count booklets are available for Logan Airport employees who passes or subway passes as an alternative to parking are doing so in don't find the pass to be economical. Employees using the service recognition that alternative modes may offer an equivalent or may park free of charge in the Logan Express parking lots. Taking shorter commute time compared to parking at the remote garage. advantage of the price incentive, employees of one airline con­ The remote garage is located in a town that has a high concen­ vinced their employer to subsidize their Logan Express passes in tration of Logan employees. Since it is within walking distance of exchange for their parking privileges. Several other airlines are now some residential areas, some employees may also use the bus as pri­ preparing to offer employees the option of a Logan Express pass or mary transportation to Logan. a subway pass in exchange for parking. The number of employees using the Logan Express compares favorably to the concentration of employees in each of the market Incentives for Massport Employees areas. In all, about 10 percent of employees reside in towns served by the three Logan Express routes. The most mature of the three Massport employees are subsidized for 50 percent of the monthly routes captures about 25 percent of its employee market on an aver­ cost of commuting by subway, bus, commuter rail, and vanpools. age weekday. On each of the routes, employees account for between The employee share of transit passes may be paid for through pay­ 5 and 11 percent of ridership. roll deduction. For Massport employees commuting by Logan 80 TRANSPORTATION RESEARCH RECORD 1506

Express or water shuttle, the 50-percent subsidy is applied to the cost of common and different employee needs. Employers would be of the already discounted monthly pass or 10-ride ticket booklet. encouraged to determine individual employee needs through sur­ veys or focus groups, and to share the results with the TMA. Initia­ tives and funding mechanisms could be developed collectively, PREVIEW OF FUTURE PROGRAM through the TMA, or by individual employers. DEVELOPMENT A TMA would be an appropriate forum for encouraging employer subsidization Of alternatives to single-occupant driving, In the upcoming months the following ideas will be further devel­ and for encouraging flextime for employees when possible. oped and evaluated, and some will be incorporated into a formal Through a TMA, a guaranteed-ride-home program could be con­ Logan employee commute options program. Because development sidered back-up transportation for commuters using alternatives to of the formal program is in its early stages and is subject to discus­ the single-occupant automobile. The guaranteed ride home would sion and collaboration with Logan Airport employers and within probably be available for employees in the case of an emergency or Massport, it is too early to provide great detail about the individual unscheduled overtime. elements or cost of the program. Additional potential elements that are likely to be studied by the TMA include the following: if potential passenger and em­ ployee demand is sufficient, working with some of the private bus Ridematching Assistance operators to offer a limited amount of nonstop trips to Logan; encouraging private carriers to offer deeper discounts; and adding Current employee work hours and geographic considerations limit links to existing services. The level of ridership needed to support the potential for capturing large concentrations of employees in direct trips will vary by carrier due to different operating costs and existing scheduled services. For the same reasons, the outlook for revenues. Work with the private carriers on direct trips and fare cost-effective new services is not good. In 1992 Massport con­ reductions is more likely to be successful through a TMA, since ducted a survey of employees residing in a town with one of the operators have been skeptical of employee demand when highest concentrations of employees. Because the town does not approached by Massport about potential fare reductions. have convenient access to public transportation, Massport consid­ ered initiating a shuttle bus network similar to a school bus network. Survey results indicated that, due to dispersed employee hours and Air Passenger Services residences, such a service would not be financially feasible. Given the above conditions, along with current employment lev­ Massport is continuously exploring the potential for additional air els, carpooling may be a more realistic alternative for many airport passenger services. Any new service would capture some employ­ employees. Massport is considering acquisition of a computerized ees. Studies are currently under way for another Logan Express ser­ ridematching program that enables employees to enter personal vice and for alternative transportation options for passengers whose commute information directly into a data base by telephone, for trips originate in high-density, close-in communities. The needs of temporary or part-time matching, _or to become a permanent mem­ these passengers would not be met by a traditional scheduled bus ber of a carpool or vanpool. The program does not require person­ service. Massport may also provide a nonstop bus link between the nel to assist in the matching. This would allow an employee to call airport and a downtown intermodal facility upon opening of the new a dedicated telephone number and communicate personal commute cross-harbor tunnel in early 1996. The intermodal facility, called information using the telephone key pad. The information would be South Station, is located about three kilometers from Logan Airport, instantly processed, and the employee would be provided with near the financial district. South Station serves as a collection point ridesharing and high-occupancy vehicle alternatives. This would for all commuter rail lines south of Boston, intercity rail, and some enable an employee with a varying schedule to participate in flexi­ public and private bus routes. A high-frequency bus link between ble carpooling, that is, as a part-time or temporary member of sev­ Logan and South Station will eliminate two transfers for both sub­ eral carpools. Mass port will also explore the potential for on-airport way and commuter rail passengers, providing a better level of ser­ priority parking for those who choose ridesharing over single­ vice for all Logan users. occupant automobile access.

CONCLUSION Logan Airport Transportation Management Association Major airports like Logan are large traffic generators, and are viewed by many as an easy target for trip reduction. But the trip­ Massport will probably provide some start-up funds for the forma­ reduction strategies that may influence airport employees are dif­ tion of a transportation management association (TMA) among ferent from what is thought of for typical commuters. Logan employers. The TMA can then study and recommend addi­ Airport employees are not as influenced by trip-reduction strate­ tional elements for the employee commute options program. Mass­ gies as typical commuters because of their special scheduling needs, port believes that a program developed under a TMA will be more the availability of fully subsidized on-airport or near-airport park­ successful than a program developed by Massport, as participation ing, and other considerations specific to an individual airport. in a TMA would demonstrate the employers' support of the com­ Because the highest employee-origin densities in the greater mute options program. It would also allow an exchange of ideas Boston area are not in towns with a high density of air passengers, among employers, including the insights of national companies who travel mode choices available to air passengers are not viable are dealing with employee commute issues in a variety of cities options for many Logan employees. Because employee work hours across the country. The TMA would facilitate the communication are dispersed over all hours of the day and week, it is currently not Ricard 81

cost-effective for Mass port to develop dedicated employee services, ACKNOWLEDGMENTS even for towns with a high density of employees. A successful airport employee commute program will offer a The author expresses her gratitude to John Passaro for his editing variety of options to meet employee commuting needs. Flexible car­ assistance. pools may be a realistic alternative for many employees. Lower­ The 1990 Employee Survey was conducted by Charles River cost initiatives that may accommodate some employees, such as Associates, under contract to Massport. flextime, where applicable, and a guaranteed ride home program, will make carpooling or vanpooling more attractive to employees. Where feasible, existing high-occupancy modes may be made more accessible to employees by adding trips, offering reduced REFERENCES fares, or adding a limited amount of direct service on select routes. Alternatives to the single-occupant automobile will become more 1. Addante, E. and D. Ricard. Ground Access Strategies and Alternatives at attractive for employees if the supply of on-airport parking is re­ Boston Logan International Airport. Proceedings of Seminar B, Airport Planning Issues, 22nd European Transport Forum, Volume P374, PTRC duced, particularly if on-airport priority parking is made available Education and Research Services Ltd., London, England, September for carpools and vanpools. Alternatives to the single-occupant 1994, pp. 27-36. automobile will also become more attractive if employees be­ 2. Federal Aviation Administration, DOT. Flight Time Limitations and come responsible for parking costs; however, this is unlikely at Rest Requirements: Domestic Air Carriers. 14 CFR 121.471, Subpart Q, Code of Federal Regulations, Aeronautics and Space. Revised as of Jan­ this time, due to collective bargaining agreements, nationwide air­ uary 1, 1994. line employee benefits packages, and competitive employee bene­ fits among airlines. Offering employees the cash equivalent of park­ ing fees to be used for parking or for a less expensive alternative mode may also influence employees away from the single-occupant Publication of this paper sponsored by Committee on Airport Terminals and commute. Ground Access.